[UNTITLED] Oxford Handbooks Online [UNTITLED] The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Subject: Economics and Finance Online Publication Date: Nov 2012 (p. iv) Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide. Oxford New York Auckland Cape Town Dar es Salaam Hong Kong Karachi Kuala Lumpur Madrid Melbourne Mexico City Nairobi New Delhi Shanghai Taipei Toronto With offices in Argentina Austria Brazil Chile Czech Republic France Greece Guatemala Hungary Italy Japan Poland Portugal Singapore South Korea Switzerland Thailand Turkey Ukraine Vietnam Oxford is a registered trademark of Oxford University Press in the UK and certain other countries. Published in the United States of America by Oxford University Press 198 Madison Avenue, New York, NY 10016 © Oxford University Press 2012 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, without the prior permission in writing of Oxford University Press, or as expressly permitted by law, by license, or under terms agreed with the appropriate reproduction rights organization. Inquiries concerning reproduction outside the scope of the above should be sent to the Rights Department, Oxford University Press, at the address above. You must not circulate this work in any other form and you must impose this same condition on any acquirer. Page 1 of 2 [UNTITLED] Library of Congress Cataloging-in-Publication Data The Oxford handbook of the digital economy / edited by Martin Peitz and Joel Waldfogel. p. cm. Includes bibliographical references and index. ISBN 978–0–19–539784–0 (cloth : alk. paper) 1. Information technology— Economic aspects—Handbooks, manuals, etc. 2. Electronic commerce— Handbooks, manuals, etc. 3. Business enterprises—Technological innovations—Handbooks, manuals, etc. I. Peitz, Martin. II. Waldfogel, Joel, 1962– III. Title: Handbook of the digital economy. HC79.I55O87 2012 384.3′3—dc23 2012004779 ISBN 978–0–19–539784–0 1 3 5 7 9 8 6 4 2 on acid-free paper Page 2 of 2 Contents Front Matter Consulting Editors [UNTITLED] Contributors Introduction Infrastructure, Standards, and Platforms Internet Infrastructure Shane Greenstein Four Paths to Compatibility Joseph Farrell and Timothy Simcoe Software Platforms Andrei Hagiu Home Videogame Platforms Robin S. Lee Digitization of Retail Payments Wilko Bolt and Sujit Chakravorti Mobile Telephony Steffen Hoernig and Tommaso Valletti Two­Sided B to B Platforms Bruno Jullien The Transformation of Selling Online versus Offline Competition Ethan Lieber and Chad Syverson Comparison Sites José­Luis Moraga­González and Matthijs R. Wildenbeest Price Discrimination in the Digital Economy Drew Fudenberg and J. Miguel Villas­Boas Bundling Information Goods Jay Pil Choi Internet Auctions Ben Greiner, Axel Ockenfels, and Abdolkarim Sadrieh Reputation on the Internet Luís Cabral Advertising on the Internet Simon P. Anderson User­Generated Content Go to page: Incentive­Centered Design for User­Contributed Content Lian Jian and Jeffrey K. Mackie­Mason Social Networks on the Web Sanjeev Goyal Open Source Software Justin P. Johnson Threats Arising from Digitization and the Internet Digital Piracy: Theory Paul Belleflamme and Martin Peitz Digital Piracy: Empirics Joel Waldfogel The Economics of Privacy Laura Brandimarte and Alessandro Acquisti Internet Security Tyler Moore and Ross Anderson End Matter Index Contributors Oxford Handbooks Online Contributors The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Subject: Economics and Finance Online Publication Date: Nov 2012 Contributors Alessandro Acquisti is Associate Professor of Information Technology and Public Policy at Heinz College at Carnegie Mellon University. Ross Anderson is Professor in Security Engineering at the University of Cambridge Computer Laboratory. Simon P. Anderson is Commonwealth Professor of Economics at the University of Virginia. Paul Belleflamme is Professor of Economics at Université Catholique de Louvain. Wilko Bolt is an economist at the Economics and Research Division of De Nederlandsche Bank. Laura Brandimarte is a PhD candidate in the Public Policy Department (Heinz College) at Carnegie Mellon University. Sujit “Bob” Chakravorti is the Chief Economist and Director of Quantitative Analysis at The Clearing House. Page 1 of 4 Contributors Luis Cabral is Professor of Economics and Academic Director (New York Center), IESE Business School, University of Navarra and W.R. Berkley Term Professor of Economics at the Stern School of Business at New York University. Jay Pil Choi is Scientia Professor in the School of Economics at the Australian School of Business, University of New South Wales. Joseph Farrell is Professor of Economics at the University of California–Berkeley. Drew Fudenberg is the Frederick E. Abbe Professor of Economics at Harvard University. Sanjeev Goyal is Professor of Economics at the University of Cambridge. Shane Greenstein is the Elinor and H. Wendell Hobbs Professor of Management and Strategy at the Kellogg School of Management at Northwestern University. Ben Greiner is Lecturer at the School of Economics at the University of New South Wales. Andrei Hagiu is Assistant Professor of Strategy at the Harvard Business School. (p. viii) Steffen Hoernig is Associate Professor with “Agregação” at the Nova School of Business and Economics in Lisbon. Lian Jian is Assistant Professor of Communication and Journalism at the Annenberg School, University of Southern California. Justin Johnson is Associate Professor of Economics at the Johnson Graduate School of Management at Cornell University. Page 2 of 4 Contributors Bruno Jullien is a Member of the Toulouse School of Economics and Research Director at CNRS (National Center for Scientific Research). Robin S. Lee is Assistant Professor of Economics at the Stern School of Business at New York University. Ethan Lieber is a PhD student in the Economics Department at the University of Chicago. Jeffrey K. MacKie-Mason is the Dean of the School of Information at the University of Michigan. Jose-Luis Moraga-Gonzalez is Professor of Microeconomics at VU University– Amsterdam and Professor of Industrial Organization at the University of Groningen. Tyler Moore is a postdoctoral fellow at the Center for Research on Computation and Society at Harvard University. Axel Ockenfels is Professor of Economics at the University of Cologne. Martin Peitz is Professor of Economics at the University of Mannheim. Abdolkarim Sadrieh is Professor of Economics and Management at the University of Magdeburg. Timothy Simcoe is Assistant Professor of Strategy and Innovation at the Boston University School of Management. Chad Syverson is Professor of Economics at the Booth School of Business, University of Chicago. Page 3 of 4 Contributors Tommaso Valletti is Professor of Economics at the Business School at Imperial College, London. J. Miguel Villas-Boas is the J. Gary Shansby Professor of Marketing Strategy at the Haas School of Business, University of California–Berkeley. Joel Waldfogel is the Frederick R. Kappel Chair in Applied Economics at the Carlson School of Management at the University of Minnesota and a Research Associate at the National Bureau of Economic Research. Matthijs R. Wildenbeest is Assistant Professor of Business Economics and Public Policy at the Kelley School of Business, Indiana University. Page 4 of 4 Introduction Oxford Handbooks Online Introduction The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Subject: Economics and Finance Online Publication Date: Nov 2012 Introduction Digitization—and the Internet—have transformed business and society in the past 15 years. The world's Internet-connected population has grown from essentially zero in 1995 to 2 billion today.1 The Internet and digitization have transformed many industries, including retailing, media, and entertainment products. Millions of people spend hours each day accessing and creating information online. The firms of the digital economy not only affect the daily life of most people in industrialized countries, but they are also highly profitable. Once nearly left for dead, Apple was in 2011 the third most valuable company in the world, with a market valuation of almost $300 billion US, far above its new-economy forebears Microsoft, IBM, and AT&T at $239, $182, and $174, respectively. Rounding out the newer generation of digital firms are Google, valued at $192 billion, Amazon at $77 billion, Facebook at $83 billion, and eBay at $36 billion.2 While particular valuations fluctuate, it is clear that firms actively in the digital economy are now global players with market capitalizations exceeding those of long-established companies in more traditional industries. New digitally enabled technologies have facilitated more extensive application of many business strategies that had formerly occupied the attention of economic theorists more than business practitioners. The practices include auctions, price discrimination, and product bundling. Although they have been employed for centuries, they are easier to utilize in digital contexts. Software operating on platforms has heightened the role of platform competition and network effects. The Internet has also provided a stage for many phenomena that would have been difficult for economists—or others—to imagine a few decades ago: large-scale sharing of digital material at sites such as YouTube, Facebook, and Wikipedia. Open-source software reflects related behaviors. Along with many new opportunities, the Internet has also brought some new threats. These include threats to businesses, such as piracy and the security of connected devices, as well as threats to individuals, such as privacy concerns. Page 1 of 4 Introduction These developments have prompted what is, by academic standards, a rapid outpouring of research. This volume is an attempt to describe that work to date, (p. x) with the goals of both explicating the state of the literature and pointing the way toward fruitful directions for future research. The book's chapters are presented in four sections corresponding four broad themes: (1) infrastructure, standards, and platforms; (2) the transformation of selling, encompassing both the transformation of traditional selling and new, widespread application of tools such as auctions; (3) user-generated content; and (4) threats in the new digital environment. No chapter is a prerequisite for another, and readers are encouraged to read in any order that aids digestion of the material. A guide to the chapters follows. The first section deals with infrastructure, standards, and platform competition. In chapter 1, Shane Greenstein provides an overview on Internet infrastructure with a particular emphasis on Internet access and broadband development. There is an interdependence between infrastructure development and changes in industries that rely on this infrastructure. Thus, researchers investigating such industries need a proper understanding of Internet infrastructures. In chapter 2, Joseph Farrell and Timothy Simcoe discuss the economics of standardization, and describe the costs and benefits of alternative ways to achieve compatibility. The section then turns to issues related to platform competition. The next chapters focus on a number of industries that heavily rely on recent developments in electronic data storage and transmission. In chapter 3, Andrei Hagiu provides a detailed view on developments in business strategies in the market of software platforms. He offers some rich institutional details on this important and rapidly changing industry. In chapter 4, Robin Lee provides a related analysis for the videogame industry. In chapter 5, Wilko Bolt and Sujit “Bob” Chakravorti survey the research on electronic payment systems. In chapter 6, Steffen Hoernig and Tomasso Valletti offer an overview on recent advances on the economics of mobile telephony. All these industries are characterized by mostly indirect network effects. Here, to understand the development of these industries, one has to understand the price and non-price strategies chosen by platform operators. Thus, an important part of the literature relates to the recent theory of two-sided markets. This theme will reappear in a number of other chapters in this handbook. Finally, in chapter 7, Bruno Jullien contributes to our understanding of B2B platforms, drawing on insights from the two-sided market literature. This theory-oriented chapter also helps in formalizing other platform markets. The second section deals with the transformation of selling. The Internet has transformed selling in a variety of ways: the reduced costs of online retailing threatens offline retailers, widespread availability of information affects competition, digital technology allows the widespread employment of novel pricing strategies (bundling, price discrimination), and auctions are now seeing wide use. In chapter 8, Chad Syverson and Ethan Lieber focus on the interdependence between online and offline retailing. In chapter 9, Jose-Luis Moraga and Matthijs Wildenbeest survey the work on comparison sites. They develop a formal model that reproduces some of the insights of the literature. Within the general topic of selling, the volume then turns to pricing practices. In chapter 10, Drew Fudenberg and Miguel Villas-Boas start with the observation (p. xi) that due to Page 2 of 4 Introduction advances in information technology, firms have ever more detailed information about their prospective and previous customers. In particular, when firms have information about consumers’ past purchase decisions, they may use this information for price discrimination purposes. The focus of the chapter is on the effects of price discrimination that is based on more detailed customer information, both under monopoly and under competition. In chapter 11, Jay Pil Choi surveys the literature on product bundling. This survey focuses on the theory of product bundling and to which extent it is particularly relevant in the context of the digital economy. The next two chapters address issues related to auctions. In chapter 12, Ben Greiner, Axel Ockenfels, and Abdolkarim Sadrieh provide an overview on Internet auctions. They review theoretical, empirical, and experimental contributions that address, in particular, bidding behavior in such auctions and the auction design in single-unit Internet auctions. In chapter 13, Luis Cabral surveys recent, mostly empirical work on reputation on the Internet, with a particular focus on eBay's reputation system. The selling section concludes with a contribution on advertising. In chapter 14, Simon Anderson reports on recent developments in online advertising and offers a mostly theoryoriented survey on advertising on the Internet. In particular, he develops a logit model to address some of the most important issues, which draws on the two-sided market literature. The third section of the book discusses the emergent phenomenon of user generated content on the Internet. In chapter 15, Lian Jian and Jeff MacKie Mason elaborate on user-generated content, an issue that had hardly arisen prior to the digital economy. In chapter 16, Sanjeev Goyal discusses the importance of the theory of social network for our understanding of certain features of the digital economy, in particular, the functioning of social networking sites such as Facebook. In contrast with the literature that postulates network effects in the aggregate, the literature on social networks takes the concrete structure of the network seriously and is concerned in particular with local interactions. In chapter 17, Justin Johnson elaborates on the economics of open source. Partly, open source is user-generated content. An important question concerns the generation of ideas, mostly as software products, with respect to the advantages and disadvantages of open source compared to traditional proprietary solutions. In this context, the openness of a product is also discussed. The fourth section of the volume discusses threats arising from digitization and the Internet. Chapters 18 and 19 analyze digital piracy. In chapter 18, Paul Belleflamme and Martin Peitz survey the theoretical literature on digital piracy, whereas in chapter 19, Joel Waldfogel surveys the empirical literature. In chapter 20, Alessandro Acquisti and Laura Brandimarte introduce the important issue of privacy in the digital economy. In chapter 21, Ross Anderson and Tyler Moore survey the work on Internet security. (p. xii) Notes Notes: Page 3 of 4 Introduction (1.) See Internet World Stats, at http://www.internetworldstats.com/stats.htm, last accessed June 10, 2011. (2.) See Ari Levy on http://www.bloomberg.com/news/2011-01-28/facebook-s-82-9-billionvaluation-tops-amazon-com-update1-.html, last accessed June 2, 2011. Page 4 of 4 Internet Infrastructure Oxford Handbooks Online Internet Infrastructure Shane Greenstein The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0001 Abstract and Keywords This article presents an overview on Internet infrastructure, highlighting Internet access and broadband development. Internet infrastructure is a collective term for all the equipment, personnel, and organizations that support the operation of the Internet. The Internet largely used existing capital at many installations with comparatively minor software retrofits, repurposing capital for any new function as long as it remained consistent with end-to-end. As Internet applications became more popular and more widely adopted, users pushed against the bandwidth limits of dial-up Internet access. Broadband gave users a better experience than dial-up access. The most notable feature of the governance of Internet infrastructure are the differences with the governance found in any other communications market. Many key events in Internet infrastructure took place within the United States, but this seems to be less likely as the commercial Internet grows large and more widespread. Keywords: Internet infrastructure, Internet access, broadband, United States, governance 1. Introduction Internet infrastructure is a collective term for all the equipment, personnel, and organizations that support the operation of the Internet. The commercial Internet arose after many firms and users voluntarily adopted a set of practices for enabling internetworking, namely, transferring data between local area networks and computer clients. The commercial Internet began to provide many revenue-generating services in the mid-1990s. As of this writing, this network supports a wide array of economic services to more than a billion users, and it continues to grow worldwide. Generally speaking, four types of rather different uses share the same Internet infrastructure: browsing and e-mail, which tend to employ low bandwidth and which can tolerate delay; video downloading, which can employ high bandwidth and can tolerate some delay; voice-over Internet protocol (IP) and video-talk, which tend to employ high bandwidth and whose quality declines with delay; and peer-to-peer applications, which tend to use high bandwidth for sustained periods and can tolerate delay, but, in some applications (such as Bit-Torrent), can impose delay on others.1 The precise economic characteristic of Internet infrastructure defies simple description. Like other private assets, sometimes Internet infrastructure is an input in the production of a pecuniary good, and regular investment extends the functionality or delays obsolescence. Like a public good, sometimes more than one user employs Internet infrastructure without displacing another or shaping the (p. 4) quality of another user's experience. Even when visible, the economic contribution of Internet infrastructure is often measured indirectly at best. Yet in contrast to many private assets or public goods, sometimes many decision makers—instead of one—govern the creation and deployment of Internet infrastructure. Moreover, although market-oriented prices influence the investment decisions of some participants, something other than market prices can shape the extent of investment and the Page 1 of 24 Internet Infrastructure use of such investment. This array of applications, the mix of economic characteristics, and the economic stakes of the outcomes have attracted considerable attention from economic analysts. This review summarizes some of the key economic insights of the sprawling and vast literature on Internet infrastructure. The chapter first provides a summary of the emergence of the commercial Internet, providing a review of the common explanations for why its structure and pricing took a specific form. The chapter then describes how the deployment of broadband access and wireless access altered many aspects of Internet infrastructure. It also reviews the importance of platforms, another new development that has changed the conduct of firms providing infrastructure for the commercial Internet. The chapter finishes with a review of the private governance of Internet infrastructure and the role of key government policies. 2. The Emergence of the Commercial Internet Which came first, the infrastructure or the commercial services? In the popular imagination the Internet emerged overnight in the mid-1990s, with the creation of the commercial browser and services employing that browser. In fact, the Internet developed for more than two decades prior to the emergence of the browser, starting from the first government funding at the Defense Advanced Research Project Agency (DARPA) in the late 1960s and the National Science Foundation (NSF) in the 1980s.2 Hence, there is no question that the infrastructure came first. The Internet infrastructure of the mid-1990s employed many components and inventions inherited from the government-funded efforts. Both the commercial Internet and its government-sponsored predecessor are packet switching networks.3 While the Internet is certainly not the first packet switching network to be deployed, the commercial Internet has obtained a size that makes it the largest ever built. In a packet switching network a computer at one end packages information into a series of discrete messages. Each message is of a finite size. As part of initial processing by the packet switching system, larger messages were divided into smaller packets. All of the participating computers used the same conventions for messages and packets. Messages successfully traveled between computers when the (p. 5) sending and receiving computer standardized on the same procedures for breaking up and reassembling packets. Sending data between two such locations rely on something called “protocols,” standardized software commands that organize the procedures for moving data between routers, computers, and the various physical layers of the network. Protocols also define rules for how data are formatted as they travel over the network. In 1982 one design became the standard for the network organized by DARPA, a protocol known as TCP/IP, which stands for transmission control protocol/internet protocol. Vint Cerf and Robert Kahn wrote the first version, replacing an outmoded protocol that previously had run the DARPA network. Over time the technical community found incremental ways to change the protocol to accommodate large-scale deployment of infrastructure and applications.4 TCP/IP had the same role in the government-sponsored and commercial Internet. It defined the “headers” or labels at the beginning and end of a packet.5 Those headers informed a computer processor how to reassemble the packets, reproducing what had been sent. By the mid-1980s all Unix-based computing systems built that interface into their operating systems, and in the 1990s virtually every computing system could accommodate it.6 An example of TCP/IP in action may help to illustrate its role. Consider sending messages from one computer, say, at Columbia University in New York, to another, say, at Stanford University in California. In the 1980s such an action involved the following steps. A user of a workstation, typically using a computer costing tens of thousands of dollars, would compose the message and then issue a command to send it. This command would cause the data to be broken into TCP/IP packets going to a central mail machine at Columbia and then a destination mail machine at Stanford. In between the central mail machine and destination mail machine the file would be broken down into pieces and completely reassembled at the final destination. During that transmission from one machine to another, the file typically traveled through one or more routers, connecting various networks and telephone lines. At the outbound site it was typically convenient to send through a central mailserver, though that did not need to be true Page 2 of 24 Internet Infrastructure at the inbound site. At the Stanford server, the local computer would translate the data into a message readable on another person's workstation. If something went wrong—say, a typographical error in the address—then another set of processes would be initiated, which would send another message back to the original author that something had failed along the way.7 These technical characteristics define one of the striking economic traits of Internet infrastructure: it requires considerable coordination. First, designers of hardware and software must design equipment to work with other equipment, and upgrades must meet a similar requirement. Second, efficient daily operations of the Internet require a wide array of equipment to seamlessly operate with one another. Indeed, the widespread adoption of TCP/IP partly explains how so many participants coordinate their activity, as the protocol acts as a focal point around which many participants organize their activities at very low transaction costs. (p. 6) Another key factor, especially during this early period, helped coordinate activity, a principle called end-to-end.8 This principle emerged in the early 1980s to describe a network where the switches and routers retained general functionality, moving data between computers, but did not perform any processing. This was summarized in the phrase, “The intelligence resided at the edges of the network.” Any application could work with any other at any edge, so long as the routers moved the data between locations.9 End-to-end had substantial economic consequences for the deployment of the commercial Internet. At the time of its invention, it was a radical engineering concept. It differed from the principles governing the telephone network, where the switching equipment performed the essential processing activities. In traditional telephony, the end of the network (the telephone handset) contained little functionality.10 For a number of reasons, this departure from precedent was supported by many outsiders to the established industry, especially researchers in computing. Hence, the Internet encountered a mix of benign indifference and active resistance from many established firms in the communications market, especially in the telephone industry, and many did not invest in the technology, which later led many to be unprepared as suppliers to the commercial Internet in the mid-1990s.11 End-to-end also had large economic consequences for users. First, the Internet largely employed existing capital at many installations with comparatively minor software retrofits, repurposing capital for any new function as long as it remained consistent with end-to-end. Repurposing of existing capital had pragmatic economic advantages, such as lowering adoption costs. It also permitted Internet applications and infrastructure to benefit from the same technical changes shaping other parts of computing.12 Accommodating heterogeneous installations supported another—and, in the long run, especially important—economic benefit: It permitted users with distinct conditions to participate on the same communications network and use its applications, such as email. 3. The Structure of the Commercial Internet As long as the Internet remained a government-managed enterprise, the NSF-sponsored Internet could carry only traffic from the research community, not for commercial purposes.13 In the late 1980s the NSF decided to privatize the operations of a piece of the Internet it managed (while the military did not privatize its piece). One key motivation for initiating this action was achievement of economies of scale, namely, the reduction in costs that administrators anticipated would emerge if researchers shared infrastructure with private users, and if competitive processes drove private firms to innovate.14 The transition reached resolution by the mid-1990s. Many of the events that followed determined the structure of (p. 7) supply of Internet infrastructure for the next decade and a half, particularly in the United States.15 The first group of backbone providers in the United States (i.e., MCI, Sprint, UUNET, BBN) had been the largest carriers of data in the NSF network. In 1995 and 1996, any regional Internet service provider (ISP) could exchange traffic with them. At that time the backbone of the US Internet resembled a mesh, with every large firm both interconnecting with every other and exchanging traffic with smaller firms.16 Some of these firms owned their own fiber (e.g., MCI) and some of them ran their backbones on fiber rented from others (e.g., UUNet). After 1997 a different structure began to take shape. It was partly a mesh and partly hierarchical, using “tiers” to describe a hierarchy of suppliers.17 Tier 1 providers were national providers of backbone services and charged a fee to smaller firms to interconnect. The small firms were typically ISPs that ranged in size and scale from wholesale regional firms down to the local ISP handling a small number of dial-in customers. Tier 1 firms did most of what became known as “transit” data services, passing data from one ISP to another ISP, or passing from a content firm Page 3 of 24 Internet Infrastructure to a user. In general, money flowed from customers to ISPs, who treated their interconnection fees with backbone firms as a cost of doing business. Tier 1 firms adopted a practice known as “peering,” and it appeared to reinforce the hierarchical structure. Peering involved the removal of all monetary transfers at a point where two tier 1 providers exchanged traffic. Peering acknowledged the fruitlessness of exchanging money for bilateral data traffic flows of nearly equal magnitude at a peering point. Hence, it lowered transaction costs for the parties involved. However, because their location and features were endogenous, and large firms that denied peering to smaller firms would demand payment instead, many factors shaped negotiations. As a result, the practice became controversial. An open and still unresolved policy-relevant economic question is whether peering merely reflects a more efficient transaction for large-scale providers or reflects the market power of large suppliers, providing additional competitive advantages, which smaller firms or non–tier 1 firms cannot imitate. The economics of peering is quite challenging to characterize generally, so this question has received a range of analysis.18 Another new feature of Internet infrastructure also began to emerge at this time, third-party caching services. For example, Akamai, a caching company, would pay ISPs to locate its servers within key points of an ISP's network. Content providers and other hosting companies would then pay the caching companies to place copies of their content on such services in locations geographically close to users, aspiring to reduce delays for users. Users would be directed to the servers instead of the home site of the content provider and, thus, receive faster response to queries. A related type of service, a content delivery network (CDN), provided similar services for electronic retailers. In general, these became known as “overlays,” and since these were not part of the original design of the noncommercial Internet, numerous questions emerged about how overlays altered the incentives to preserve end-to-end principles.19 These changes also raised an open question about whether a mesh still determined most economic outcomes. For much Internet service in urban areas the (p. 8) answer appeared to be yes. ISPs had options from multiple backbone providers and multiple deliverers of transit IP services, and many ISPs multihomed to get faster services from a variety of backbone providers. ISPs also had multiple options among cache and CDN services. Evidence suggests that prices for long-distance data transmission in the United States continued to fall after the backbone privatized, reflecting excess capacity and lower installation costs.20 For Internet service outside of urban areas the answer appeared to be no. ISPs did not have many options for “middle mile” transit services, and users did not have many options for access services. The high costs of supply made it difficult to change these conditions.21 Perhaps no factor altered the structure of supply of Internet infrastructure after commercialization more than the World Wide Web: The World Wide Web emerged just as privatization began, and it is linked with a particularly important invention, specifically, the commercial browser. As the commercial Internet grew in the mid to late 1990s, traffic affiliated with the web overtook electronic mail, symptomatic of web applications as the most popular on the Internet. The growth in traffic had not been anticipated at the time NSF made plans for privatizing the backbone, and the subsequent growth in traffic fueled a global investment boom in Internet infrastructure. Tim Berners-Lee built key parts of the World Wide Web.22 In 1991 he made three inventions available on shareware sites for free downloading: html (hypertext markup language) and the URL (universal resource locator), a hyper-text language and labeling system that made transfer of textual and nontextual files easier using http (hyper-text transfer protocol). By 1994 and after the plans for commercialization were set and implemented, a new browser design emerged from the University of Illinois. It was called Mosaic, and it was widely adopted in the academic Internet. Mosaic became the model that inspired the founding of Netscape, the first successful commercial browser, which also motivated Microsoft to develop Internet Explorer, which touched off the browser wars.23 As of this writing most observers expect data traffic to continue to grow at rates in the neighborhood of 40 percent to 50 percent a year.24 This is due to the emergence of another set of applications. Data supporting peer-to-peer and video applications have been growing rapidly after the millennium, each year using larger fractions of available capacity.25 Page 4 of 24 Internet Infrastructure 4. Pricing Internet Access The value chain for Internet infrastructure begins with the pricing for Internet access. Households and business establishments pay for access to the Internet.26 ISPs provide that access and use the revenue to pay for inputs, namely, other Internet infrastructure. (p. 9) In the first half of the 1990s, most commercial ISPs tried the same pricing norms governing bulletin boards. For bulletin boards the pricing structure of the majority of services involved a subscription charge (on a monthly or yearly basis) and an hourly fee for usage. For many applications, users could get online for “bursts” of time, which would reduce the size of usage fees. The emergence of faster and cheaper modems in the mid-1990s and largescale modem banks with lower per-port costs opened the possibility for a different pricing norm, one that did not minimize the time users spent on the telephone communicating with a server. The emergence of low-cost routines for accessing a massive number of phone lines also contributed to a new norm, because it enabled many ISPs to set up modem banks at a scale only rarely seen during the bulletinboard era. As the Internet commercialized, two broad viewpoints emerged about pricing. One viewpoint paid close attention to user behavior. Users found it challenging to monitor time online, often due to multiple users within one household. The ISPs with sympathy for these user complaints priced their services as unlimited usage for a fixed monthly price. These plans were commonly referred to as flat-rate or unlimited plans. Another viewpoint regarded user complaints as transitory. Supporters of this view pointed to cellular telephones and bulletin boards as examples where users grew accustomed to pricing-per-minute. The most vocal supporter for this view was one up-andcoming bulletin-board firm, America Online (now known simply as AOL), which had seemed to grow with such usage pricing. As it turned out, flat rate emerged as the dominant pricing mode for wireline access. There were many reasons for this. For one, the US telephone universal service policy had long subsidized local landline household calls by charging greater prices for business and long-distance calls than for local calls, which were priced at a flat rate per month in almost every state, with the definition of local left to a state regulator. Typically local was defined over a radius of ten to fifteen miles, sometimes less in dense urban areas, such as Manhattan. Thus, using local rates reduced household user expenses, thereby making the service more attractive. That enabled an ISP to offer service to nontechnical users, betting the users would not stay online for long—in effect, anywhere from twenty to thirty hours of time a month at most. Because such light users did not all dial-in on the same day or at the same time of day, the equipment investment did not need to handle all calls at once. With such light users, an ISP could serve a local area with modem bank capacity at anywhere from one-third to one-quarter the size of the total local service base. A large enough user community could thus share equipment, defraying equipment costs for an ISP offering flat rate. Experiments showed that the economies of scale for defraying equipment costs in this way could support approximately 1000 users in a local point of presence. For many dial-up ISPs, this was not a binding constraint.27 Flat rate succeeded for another reason. Other firms entered into the production of web services, complementing what ISPs provided. Flat rate appealed to users who could then choose among a wide array of media and services.28 (p. 10) Throughout 1996, 1997, and 1998, ISPs experimented with hourly limits and penalties for exceeding these caps, yet most users resisted them. The experiments were motivated by many users who stayed online for far more than twenty to thirty hours a month, thereby ruining the key economics that allowed ISPs to share equipment across many users.29 Most such limits were not particularly binding—involving monthly limits ranging from sixty to one hundred hours. Some ISPs tried offering steep discounts for steep limits, such as $10 discounts for thirty hours a month. Yet few buyers took them, persisting with the slightly more expensive unlimited contracts, typically priced as $20 per month. In short, despite all these experiments, flat rate dominated transactions during the dial-up era of the commercial Internet.30 5. The Broadband Upgrade As Internet applications became more popular and more widely adopted, users pushed against the bandwidth limits Page 5 of 24 Internet Infrastructure of dial-up Internet access. That motivated some well-placed firms to deploy and offer broadband access as Internet access. The investment of broadband initiated a second wave of investment in Internet infrastructure after the turn of the millennium. That has been coincident with the presence of more capital deepening in business computing and in many facets of the operations to support it. The two most common forms were cable modem service and (digital subscriber line) (DSL) service. Cable modem service involved a gradual upgrade to cable plants in many locales, depending on the generation of the cable system.31 Broadband over telephone lines involved upgrades to telephone switches and lines to make it feasible to deliver DSL. Both of these choices typically supported higher bandwidth to the household than from the household —thus called asymmetric digital subscriber line (ADSL). Broadband clearly gave users a better experience than dial-up access.32 Broadband provides households with faster Internet service and thus access to better online applications. Broadband also may allow users to avoid an additional phone line for supporting dial-up. In addition, broadband services are also “always on,” and users perceive that as a more convenient service. It is also generally faster in use. A maximum rate of 14.4K (kilobytes per second) and 28.8K were predominant in the mid-1990s for dial-up modems. The typical bandwidth in the late 1990s was 43K to 51K, with a maximum of 56K. DSL and cable achieved much higher maximum bandwidths, typically somewhere in the neighborhood of a maximum rate of 750K to 3M (megabytes per second), depending on the user choices and vendor configuration. Even higher bandwidth became available to some households later. Click to view larger Figure 1.1 Percentage of Households with Computers and Internet Connections, Selected Years, 1997– 2009. Source: NTIA (2010). This story is consistent with Figure 1.1, which summarizes US government surveys of broadband use at households. The first survey questions about broadband (p. 11) use appear in 2000 and show a growth in adoption, reaching close to 20 percent of households in 2003, when these surveys were discontinued for some time.33 The survey resumed in 2007 and the anticipated trajectory continued, with 50.8 percent of households having broadband in October 2007 and 63.5 percent in October 2009. In the earliest years supply-side issues were the main determinants of Internet availability and, hence, adoption. Cable and telecom operators needed to retrofit existing plants, which constrained availability in many places. In those years, the spread of broadband service was much slower and less evenly distributed than that of dial-up service. Highly populated areas were more profitable due to economies of scale and lower last-mile expenses. As building has removed these constraints, demandrelated factors—such as price, bandwidth, and reliability—have played a more significant role in determining the margins between who adopts and who does not.34 Suppliers experimented to find appropriate structures for providing broadband access.35 Countries differed significantly in the extent to which these different delivery channels played a role. Some cable firms built out their facilities to deliver these services in the late 1990s, and many—especially telephone companies—waited until the early to mid-2000s. In some rich countries there was growing use of a third and fourth delivery channel, fiber to the home, and access with mobile modes.36 A similar wave of investment occurred in many developed countries over the first decade of the new millennium. Figure 1.2 shows the subscribers per 100 inhabitants in many countries in 2009.37 Although these numbers must be interpreted with caution, a few facts stand out. A few dozen countries in the Organisation for Economic Cooperation and Development (OECD) experienced substantial adoption of broadband, and many did not. The variance is not surprising. GDP per capita and broadband per capita have a simple correlation of 0.67.38 Page 6 of 24 Internet Infrastructure Click to view larger Figure 1.2 OECD Broadband Subscribers per 100 Inhabitants by Technology, June 2009. Source: OECD. Figure 1.3 shows the growth of subscribers per 100 inhabitants in the G7, Canada, the United States, United Kingdom, Germany, France, Italy, and Japan, (p. 12) as well as the entire OECD. Though countries differ in the level of broadband use, the similarities among them are apparent. Adoption of broadband grew in all rich and developed countries. Click to view larger Figure 1.3 Broadband Penetration, G7 Countries. Source: OECD Broadband Portal, http://www.oecd.org/sti/ict/broadband, Table 1i. To give a sense of worldwide patterns, Table 1.1 presents broadband and dial-up adoption for seven countries— the United States, Canada, United Kingdom, Spain, China, Mexico, and Brazil.39 These seven represent typical experiences in the high-income and middle-income countries of the world. The broadband data in Table 1.1 come from Point-Topic, a private consultancy.40 One fact is immediately obvious. The scale of adoption in the United States and China far outweighs the scale of adoption in any other country. That occurs for two rather obvious economic reasons. The United States and China have much larger populations than the United Kingdom, Spain, and Canada. Although Mexico and Brazil also have (p. 13) large populations, those countries had much lower rates of adoption. In short, the general level of economic development is a major determinant of adoption. Page 7 of 24 Internet Infrastructure Table 1.1 Broadband Subscribers from Point Topic Broadband Subscribers from Point Topic in Thousands Year Nation 2003 2004 2005 2006 2007 2008 2009 CAGR Brazil 634 1,442 2,671 4,278 5,691 7,509 9,480 47.2% Canada 3,706 4,829 5,809 6,982 8,001 8,860 9,528 14.4% China – 11,385 20,367 30,033 41,778 54,322 68,964 35.0% Mexico 234 429 1,060 1,945 3,106 4,774 7,836 65.1% Spain 1,401 2,524 3,444 5,469 7,322 8,296 9,023 30.5% United Kingdom 960 3,734 7,203 10,983 13,968 16,282 17,641 51.6% United States 16,042 28,770 37,576 47,489 58,791 67,536 77,334 25.2% Source: Greenstein and McDevitt (2011) The popularity of wireline broadband access has reopened questions about pricing. Following precedent, most broadband providers initially offered unlimited plans. High volume applications—such as video downloading and peer-to-peer applications—placed pressures on capacity, however. This motivated providers to reconsider their flat-rate pricing plans and impose capacity limitations. Hence, a decade and half after the blossoming of the commercial Internet there was considerable speculation about whether video applications would generate a new norm for pricing access. The answer partly depended on the regulatory regime governing Internet access and remains open as of this writing.41 6. Expanding Modes of Access In the decade after the millennium another and new set of mobile broadband services began to gain market traction with businesses and households. The first were providers known as wireless ISPs, or WISPs for short. They provided access via a variety of technologies, such as satellite, high-speed WiFi (wireless local area network that uses high-frequency radio signals to transmit and receive data), WiMax (stands for worldwide interoperability for microwave access), and other terrestrial fixed-point wireless delivery modes. These providers primarily served low-density locations where the costs of wireline access were prohibitive.42 Another wireless mode became popular, smart phones. Though smart phones had been available in a variety of models for many years, with Blackberry being among the most popular, it is commonly acknowledged that the category began (p. 14) to gain in adoption after the introduction of the Apple iPhone in 2007. As of this writing, reports suggest the Apple iPhone, Google Android, and new Blackberry designs dominate this product category for the time being. Yet competitive events remain in flux. Competitive responses organized by Palm, Microsoft, and Nokia have been attempted, and those firms and others will continue to attempt more. In short, irrespective of which firms dominate, smart phones are an important new access point. The economics of smart-phone use remain in flux as well. It is unclear whether the majority of users treat their smart phones as substitutes to their home broadband use. Smart phones provide additional mobility, and that might be a valuable trait by itself. If smart phones are simply additional services due to mobility, then the economic value of smart phone use should be interpreted one way. If the additional services are partial or complete substitutes, Page 8 of 24 Internet Infrastructure then the economic value of smart phone use should be interpreted another. Another open question concerns the boundaries between home and business use. With wireline broadband this is less of a question because the destination of the location for Internet access largely identifies its buyer (home or business). Smart phones, however, sell both to home and business, thus obliterating what had been a bright line between these kinds of customers and altering the economic analysis of the adoption decision and its impact. This also alters the potential revenue streams for Internet access, as well as the geographic distribution of use, which could lead to changes in the funding for more Internet infrastructure, as well as the geographic features of investment. In addition, because of leapfrogging of mobile over fixed broadband in emerging economies, mobile broadband may be the first broadband experience for many people. So not only is it unclear whether mobile broadband substitutes or complements fixed broadband, but the extent of substitutability could vary substantially by country according to each country's stage of infrastructure development. The changing modes of access opened multiple questions about changing norms for pricing-data services. Wireless data services over cellular networks generally did not use flat-rate pricing and often came with explicit charges for exceeding capacity limits, for example. As of this writing, the norm for these plans was, and continues to be, an open question. 7. Strategic Behavior from Suppliers When it first deployed, the Internet was called a “network of networks.” Although the phrase once had meaning as a description of the underlying configuration of infrastructure, it is misleading today. Leading firms and their business partners view the commercial Internet through the same lens they view the rest of computing. To them, the Internet is a market in which to apply their platform (p. 15) strategies. In short, the commercial Internet should be called a “network of platforms.” The list of important platforms in Internet infrastructure is long, and many involve familiar firms, such as Google, Apple, Cisco, Microsoft, and Intel. As a result, the platform strategies of private firms shape the evolution of Internet infrastructure.43 This was one of the biggest changes wrought by introducing private firms into the supply of Internet infrastructure. Commercial firms regard platforms as reconfigurable clusters and bundles of technical standards for supporting increasing functionality.44 From a user perspective, platforms usually look like “standard bundles” of components, namely, a typical arrangement of components for achieving functionality.45 Platforms matter because platform leaders compete with others. That usually leads to lower prices, competition in functionality, and accelerated rollout of new services. Many platform leaders also develop new functionality and attach it to an existing platform.46 Generally, that brings new functionality to users. A platform leader also might seek to fund businesses outside its area of expertise, if doing so increases demand for the core platform product.47 Hence, in the commercial Internet, the strategic priorities of its largest providers tend to shape the innovative characteristics of Internet infrastructure. Today many observers believe that Google—which did not even exist when the Internet first commercialized in the mid-1990s—has the most effective platform on the Internet. Hence, its behavior has enormous economic consequences. For example, sometimes it makes code accessible to programmers for mash-ups—for example, building services that attract developers and users with no direct way to generate revenue.48 Sometimes its investments have direct effects on other Internet infrastructure firms. For example, Google operates an enormous global Internet backbone, as well as CDN and caching services, for its own operations, making it one of the largest data-transit providers in the world today. The firm also uses its platform to shape the actions of others. Many other firms expend considerable resources optimizing their web pages to appear high on Google's search results, and Google encourages this in various ways. For example, it lets potential advertisers, who will bid in Google's auction, know which words are the most popular. Networking equipment provider Cisco is another prominent platform provider for Internet infrastructure, having grown to become a large provider of equipment for business establishments. For many years, Cisco made most of Page 9 of 24 Internet Infrastructure its profit from selling hubs and routers, so the platform strategy was rather straightforward. Cisco aspired to developing closely related businesses, offering users a nearly integrated solution to many networking problems. At the same time, Cisco kept out of service markets and server applications, leaving that to integrators, consultants, and software vendors. That way, Cisco did not compete with its biggest business partners. More recently, however, Cisco branched into consumer markets (with its purchase of Linksys). The firm also has moved into some server (competing with HP) and some software/service areas related to videoconferencing and telepresence (p. 16) (by purchasing Webex, for example). Cisco no longer draws the boundary where it used to, and it is unclear how wide a scope the firm wants its platform to cover. Microsoft is perhaps the next best known platform provider whose business shapes Internet infrastructure. In the early 1990s, Microsoft offered TCP/IP compatibility in Windows as a means of enhancing its networking software, as well as to support functionality in some of its applications, such as Exchange. In the mid-1990s, Microsoft offered a browser, partly as a gateway toward developing a broader array of web services, and partly for defensive purposes, to establish its proprietary standards as dominant.49 Although Microsoft continues to support these commercial positions and profit from them, the firm has not had as much success in other aspects of its commercial Internet ventures. MSN, search, mobile OS, and related activities have not yielded sustained enviable success (yet). Only the firm's investments in Xbox Live have generated a significant amount of Internet traffic, and it continues to push the boundaries of large-scale multiplayer Internet applications. Another PC firm, Intel, has an Internet platform strategy. Intel's most important Internet activity came from sponsoring a Wi-Fi standard for laptops under the Centrino brand in 2003.50 To be clear, this did not involve redesigning the Intel microprocessor, the component for which Intel is best known. It did, however, involve redesigning the motherboard for desktop PCs and notebooks by adding new parts. This redesign came with one obvious benefit: It eliminated the need for an external card for the notebook, usually supplied by a firm other than Intel and installed by users (or OEMs—original equipment manufacturers) in an expansion slot. Intel also hoped that its endorsement would increase demand for wireless capabilities within notebooks using Intel microprocessors by, among other things, reducing their weight and size while offering users simplicity and technical assurances in a standardized function. Intel hoped for additional benefits for users, such as more reliability, fewer set-up difficulties, and less frequent incompatibility in new settings. Intel has helped fund conformance-testing organizations, infrastructure development, and a whole range of efforts in wireless technology. More recently, it has invested heavily in designing and supporting other advanced wireless standards, such as WiMax. As with many other aspects of commercialization, the importance of platforms is both cause for celebration and a source of concern. It is positive when platform strategies help firms coordinate new services for users. Why is the emergence of platform providers a concern? In short, in this market structure the private incentives of large dominant firms determine the priorities for investment in Internet infrastructure. Under some circumstances dominant platform firms have incentives to deny interconnection to others, to block the expansion of others, and, in an extreme case, to smother the expansion of new functionality by others. When Internet alarmists worry about conduct of proprietary platforms, they most fear deliberate introduction of incompatibilities between platforms, and other conduct to deliberately violate end-to-end principles.51 Microsoft, AOL, Intel, Comcast, and WorldCom have all shown tendencies toward such behavior in specific episodes. (p. 17) Open-source advocates who claim that they benefit the Internet usually mean they are preventing defensive activity by leaders with defensive tendencies. More recently, a range of behavior from Apple, Facebook, Twitter, American Airlines, and the Wall Street Journal have raised fears about the “splintering” of the Internet.52 Splintering arises when users lose the ability to seamlessly move from one application to another or vendors lose a common set of platforms on which to develop their applications. Steve Jobs’ decision not to support Flash on the iPhone and iPad is one such recent example. So is Twitter's resistance to having its services searched without payment from the search engines, and Facebook has taken a similar stance. American Airlines refused to let Orbitz see its prices without modifying its website to accommodate new services American Airlines wanted to sell to frequent fliers, and Orbitz refused. As a result, the two companies no longer cooperate. For some time the Wall Street Journal has refused to let users search extensively in its archives without a subscription, and its management openly discusses aspirations to gain a fee from search engines, much as Facebook and Twitter did. What economic incentives lie behind these concerns? In a nutshell the answer begins thusly: no law compels any firm to interconnect with any other. Every firm always has the option to opt out of the Internet, and/or do something Page 10 of 24 Internet Infrastructure slightly less dramatic, such as opt out of commonly used standards, and/or try to get business partners to use proprietary standards. At any given moment, that means a firm in a strong competitive position will have incentives to raise switching costs to its installed base of developers and users, and/or deter migration of users and developers to competing platforms. A variety of strategic options might contribute to such goals, such as designing proprietary standards or blocking others from using a rival firm's standard designs. Concerns about splintering also arise because suppliers must cooperate in order to deliver services. Because one party's cost is another party's revenue, firms have incentives to sacrifice aspects of cooperation in the attempt to gain rents. Such behavior is not necessarily in users’ interests. Consider the negotiations between Cogent and Sprint, for example, which broke down after a peering dispute.53 Cogent refused to pay Sprint after Sprint insisted Congent had not met their obligations under a peering agreement. After a long stand-off, Sprint's management decided to shut down its side of the peering. That action had consequences for users on both networks who did not multihome, that is, did not use more than one backbone firm. One set of exclusive Sprint users could not reach another set of exclusive Cogent users.54 To make a long story short, users of both carriers were angry, and Sprint's management gave in after a few days. Soon after, the two firms came to a long-term agreement whose details were not disclosed publicly. Notice how inherently awkward the negotiations were: Cogent necessarily interacted or exchanged traffic with the very firm with which it competes, Sprint.55 Another set of cases illustrates how the interests of one participant may or may not intersect with the interests of all participants. For example, consider Comcast's unilateral declaration to throttle peer-to-peer (P2P) applications on its (p. 18) lines with resets.56 This case contains two economic lessons. On the one hand, it partially favors giving discretion to Comcast's management. Management could internalize the externality one user imposes on others— managing traffic for many users’ general benefit. That is, P2P applications, like Bit-Torrent, can impose large negative externalities on other users, particularly in cable architectures during peak-load time periods. Such externalities can degrade the quality of service to the majority of users without some sort of limitation or restriction. On the other hand, Comcast's behavior shapes at least one additional provider of applications, future entrepreneurs, many of whom are not present. It would be quite difficult for Comcast and future entrants to reach a bargain because some of them do not even exist yet. Eventually the FCC intervened with Comcast, issuing an order to cease blocking, which led to a court case over its authority to issue such an order. As of this writing, the full ramifications of this court case have not played themselves out. Another case involving Comcast also illustrates the open-ended nature of the cooperation between firms. In November 2010, Comcast, the largest provider of broadband Internet access in the United States, entered into a peering dispute with Level 3, one of the backbone firms with which it interconnected. Level 3 had made an arrangement with Netflix, a video provider, and this had resulted in Level 3 providing more data to Comcast than Comcast provided to Level 3. Comcast demanded that Level 3 pay for giving more data to Comcast than Comcast gave in return, to which Level 3 agreed.57 This agreement was significant because it was the first time an ISP did not pay the backbone provider for transit services, but instead, the provider paid a large ISP to reach end users. As of this writing, it is an open question how common such agreements will become, and whether they will alter the general flow of dollars among Internet infrastructure firms. 8. Governance for Internet Infrastructure A number of organizations play important roles in supporting the design, upgrading, and operations of Internet infrastructure. The most notable feature of the governance of Internet infrastructure are the differences with the governance found in any other communications market. One notable feature of this structure was the absence of much government directive or mandate. It is incorrect to say that the government was uninvolved: After all, the NSF and Department of Defense both had played a crucial role in starting and sponsoring organizations that managed and improved the operations of the Internet.58 Rather, the commercial Internet embodied the accumulation of multiple improvements suggested through a process of consensus in committees, and that consensus depended in large part on private action, what economists call (p. 19) “private orderings.”59 Unlike any other communication network, governments did not play a substantial role in these private orderings. Page 11 of 24 Internet Infrastructure The organization that governs the upgrades to TCP/IP is the Internet Engineering Task Force (IETF). It was established prior to the Internet's privatization, and continued as a nonprofit organization after its commercialization. Today it hosts meetings that lead to designs that shape the operations of every piece of equipment using TCP/IP standards.60 Many of these decisions ensured that all complying components would interoperate. Today decisions at the IETF have enormous consequences for the proprietary interests of firms. Standards committees had always played some role in the computer market, and they played a similar role in the shaping of Internet infrastructure. The Institute of Electrical and Electronics Engineers (IEEE), for example, made designs that shaped the LAN market, modem, and wireless data communications markets.61 Aside from the absence of government mandates, these groups also were notable for the absence of dominant firms. They were not beholden to the managerial auspices of AT&T or IBM, or any other large firm, for example. Though all those firms sent representatives who had a voice in shaping outcomes, these institutions were characterized by divided technical leadership. That does not imply that all these organizations conducted their business in a similar manner. On the contrary, these forums differed substantially in their conduct.62 The World Wide Web Consortium (W3C) offers an illuminating comparison. Berners-Lee forecast the need for an organization to assemble and standardize pieces of codes into a broad system of norms for operating in the hyper-text world. He founded the World Wide Web Consortium for this purpose. In 1994 he established the offices for the W3C in Cambridge, Massachusetts, just in time to support an explosion of web-based services that took advantage of existing Internet infrastructure. Berners-Lee stated that he had wanted a standardization process that worked more rapidly than the IETF but otherwise shared many of its features, such as full documentation and unrestricted use of protocols. In contrast to the IETF, the W3C would not be a bottom-up organization with independent initiatives, nor would it have unrestricted participation. Berners-Lee would act in a capacity to initiate and coordinate activities. To afford some of these, his consortium would charge companies for participating in efforts and for the right to keep up-to-date on developments.63 The governance structures for the IETF and the W3C also can be compared to what it is not, namely, the next closest alternative for global networking—the Open Systems Interconnection model, a.k.a., OSI seven-layer model. The OSI was a formal standard design for interconnecting networks that arose from an international standards body, reflecting the representation of multiple countries and participants. The processes were quite formal. The network engineering community in the United States preferred their bottom-up approach to the OSI top-down approach and, when given the opportunity, invested according to their preferences.64 The lack of government involvement could also be seen in other aspects of the Internet in the United States. For example, the Federal Communications (p. 20) Commission (FCC) refrained from mandating most Internet equipment design decisions. Just as the FCC had not mandated Ethernet design standards, so it let the spectrum become available for experiments by multiple groups who competed for wireless Ethernet standards, which eventually became Wi-Fi. Similarly, the FCC did not mandate a standard for modems other than to impose requirements that limited interference. It also did not mandate an interconnection regulatory regime for Internet carriers in the 1990s.65 The US government's most visible involvement in governance has come with its decisions for the Internet Corporation for Assigned Numbers and Names (ICANN), an organization for governing the allocation of domain names. The NSF originally took responsibility for domain names away from the academic community prior to privatizing the Internet, giving it to one private firm. The Department of Commerce established a nonprofit organization to provide oversight of this firm, with the understanding that after a decade ICANN would eventually become a part the United Nations.66 This latter transfer never took place, and, as of this writing, ICANN remains a US-based nonprofit corporation under a charter from the Commerce Department. One other notable innovative feature of Internet infrastructure is its reliance on the behavioral norms and outcomes of open-source projects. This had substantial economic consequences, establishing behavior norms for information transparency that had never before governed the design of equipment for a major communication market. Key aspects of Internet infrastructure embedded designs that emerged from designs that any firm or user could access without restriction, and to which almost any industry participant could make contributions. Page 12 of 24 Internet Infrastructure One well-known open-source project was Linux, a basis for computer operating systems. It was begun by Linus Torvald in the early 1990s as a derivative, or “fix,” to Unix. It was freely distributed, with alternating releases of a “beta” and “final” version. Starting around 1994 to 1995, about the same time as the commercialization of the Internet, Linux began to become quite popular. What had started as a small project caught the attention of many Unix users, who started contributing back to the effort. Many users began converting from proprietary versions of Unix (often sold by the hardware manufacturers) and began basing their operating systems on Linux, which was not proprietary. This movement gained so much momentum that Linux-based systems became the most common server software, especially for Internet servers.67 Apache was another early project founded to support and create “fixes” for the HTTP web server originally written by programmers at the National Center for Super Computing Applications (NCSA) at the University of Illinois. By 2006, more than 65 percent of websites in the world were powered by the Apache HTTP web server.68 Apache differed from many other open-source organizations in that contributors “earned” the right to access the code. To be a contributor one had to be working on at least one of Apache's projects. By 2006, the organization had an active and large contributor base. (p. 21) Perhaps the most well-known open source format was the least technical. It originated from something called a wiki. Developed in 1995 by Ward Cunningham, a software engineer from Portland, Oregon, wikis can either be used in a closed work group or used by everyone on the Internet. They originally were formed to replicate or make a variation on existing products or services, with the purpose of fixing bugs within the various systems. Accordingly, wikis were first developed and intended for software development but had grown out of that first use and became applied to a multitude of applications. In short, wikis became the essential software infrastructure upon which virtually all major Internet media applications are built. A particular popular application of wikis, Wikipedia, garnered worldwide attention. In the case of Wikipedia, the format was applied to the development of textual and nontextual content displayed on the Web. It is an online-only encyclopedia. The content is user-created and edited. As its homepage proudly states, it is “The Free Encyclopedia That Anyone Can Edit.” The site has always been free of charge and never accepted advertising.69 Wikipedia beat out Microsoft's Encarta for the honor of the Internet's top research site in 2005, a position that it has held ever since.70 An experimental form of copyright, the creative commons license, spread extraordinarily fast. This license, founded in only 2001, is used by over 30 million websites today.71 It has begun to play a prominent role in online experimentation and everyday activity. Creative commons licenses help organizations accumulate information in a wide array of new business formats. Flickr is one successful example, having recently passed the milestone of four billion photos on its site. The creative commons license also is employed by numerous Web2.0 initiatives and new media, which support online advertising. After so much experience with open source it is no surprise that the major participants in Internet infrastructure no longer leave these institutions alone. The standardization organizations find their committees filled with many interested participants, some with explicit commercial motives and some not. These institutions show signs of the stress, chiefly in a slowing down in their decision making, if they reach decisions at all. Perhaps that should also be cause for celebration, since it is an inevitable symptom of commercial success and the large commercial stakes for suppliers.72 9. Broadband Policy Many governments today, especially outside the United States, are considering making large subsidies for broadband investments. Some governments, such as those of South Korea and Australia, have already done so, making next-generation broadband widely available. Many years of debate in the United States led to the emergence of a National Broadband Plan, released in March 2010.73 Related debates also led to a large European framework for the governance of investments in broadband.74 (p. 22) Some of the key issues can be illustrated by events in the United States. At the outset of the commercial Internet, policy favored allowing firms to invest as they please. During the latter part of the 1990s, policy did not restrict the ability of firms to respond to exuberant and impatient demand for new Internet services. After a time, the 75 Page 13 of 24 Internet Infrastructure infrastructure to support those new services became far larger than the demand for services.75 After the dot-com boom came to a bust, the United States found itself with excessive capacity in backbone and many other infrastructure facilities to support Internet services. In the decade after, policy continued to favor privately financed investment. That resulted in a gradual building of broadband. Why did private supply favor gradualism? In short, aside from perceptions of overcapacity, few executives at infrastructure firms would ever have deliberately invested resources in an opportunity that was unlikely to generate revenue until much later, especially ten to twenty years later. Corporate boards would not have approved of it, and neither would stockholders. One of the few firms to attempt such a strategy was Verizon, which unveiled a program to build fiber to the home in the latter part of the first decade after the millennium. Due to low take-up, Verizon did not fully build these services in all its territory.76 Most arguments for building next-generation Internet broadband ahead of demand faced large political obstacles. Consider one justification, economic experimentation, namely, better broadband clearly helps experimentation in applications by making users better customers for online ads and electronic retailing. Although the monetary costs are easy to tally, the benefits are not. Relatedly, the costs are focused, but the gains are diffuse, thus making it difficult to show that broadband caused the associated gain, even if, broadly speaking, everyone recognizes that broadband raised firms’ productivity and enhanced users’ experience. Accordingly, financing broadband would involve a general tax on Yahoo and Amazon and Google and YouTube and other national electronic retailers and application providers who benefit from better broadband. Needless to say, considerable political challenges interfere with the emergence of such schemes. Some countries, such as Korea, have managed to come to such a political agreement, but these examples are the exception, not the rule. Another set of policies considers subsidizing rural broadband, that is, subsidizing the costs of building wireline broadband supply in high-cost areas.77 Since private supply already covers the least costly areas to supply, only a small fraction of potential households benefit from such subsidies, and the costs of subsidizing buildouts are especially high. Such subsidies face numerous challenges. The US National Broadband Plan provides an excellent summary of the issues. Many of the justifications for these subsidies are noneconomic in nature—aimed at increasing civic engagement among rural populations, increasing obtainment of educational goals among children, or increasing the likelihood of obtaining specific health benefits. Accordingly, the decision to provide such subsidies is often a political decision rather than purely an economic one. Another open question concerns the governance of deployment of access networks. In Europe the governments have chosen a structure that essentially (p. 23) separates ownership of transmission facilities from ownership of content, and mandates interconnection for many rivals at key points.78 In the United States there was a preference for private provision of backbone and access networks in the first decade of the millennium, and a light-handed degree of regulatory intervention, so providers were not required to offer interconnection to others. No simple statement could characterize the changing norms for regulating Internet infrastructure. For example, as of this writing, at the federal level, there are initiatives under way to adopt formal policies for ensuring the openness of Internet access.79 At the local level, there are a range of initiatives by municipalities to provide local access in competition with private suppliers.80 There is also effort to limit municipal intervention in the provision of access or deter state limitations on local initiatives.81 10. Summary No single administrative agency could possibly have built and managed the commercial network that emerged after the privatization of the Internet. The shape, speed, growth, and use of the commercial Internet after 1995 exceeded the ability of any forecaster inside or outside government circles. The value chain for Internet services underwent many changes after the Internet was privatized. More investment from private firms, and more entry from a range of overlays and new applications, altered nearly every aspect of the structure of the value chain. This evolution occurred without explicit directives from government actors, with only a light hand of directives, and with astonishing speed. Many key events in Internet infrastructure took place within the United States in the first decade and a half of the commercial Internet, but this appears to be less likely as the commercial Internet grows large and more Page 14 of 24 Internet Infrastructure widespread. While the United States continues to be the source of the largest number of users of Internet services, and the single greatest origin and destination for data traffic, the US position in the global Internet value chain will not—indeed, cannot—remain dominant. That should have enormous consequences for the evolution of the structure of global Internet infrastructure because many countries insist on building their infrastructure according to principles that differ from those that governed the first and second waves of investment in the United States. The boundaries between public and private infrastructure should change as a result, as should the characteristics of the governance and pricing of Internet infrastructure. It is no surprise, therefore, that many fruitful avenues for further economic research remain open. For example, what frameworks appropriately measure the rate of return in investment in digital infrastructure by public and private organizations? Through what mechanisms does advance Internet infrastructure produce (p. 24) economic growth, and to which industries in which locations do most of the positive and negative effects flow? What factors shape the effectiveness of different governance structures for open structures, such as those used by the IETF? What is the quantitative value of these novel governance structures? For the time being there appears to be no cessation in the never-ending nature of investment in Internet infrastructure. Indeed, as of this writing, many questions remain open about the value of different aspects of IT in the long run, and firms continue to explore approaches to creating value. Virtually all participants in these markets expect continual change, as well as its twin, the absence of economic tranquility. References Abbate, J., 1999. Inventing the Internet, MIT Press: Cambridge, Mass. Aizcorbe, A., K. Flamm, and A. Khursid, 2007. “The Role of Semiconductor Inputs in IT Hardware Price Decline: Computers vs. Communications,” in (eds.) E. R. Berndt and C. M. Hulten, Hard-to-Measure Goods and Services: Essays in Honor of Zvi Griliches, University of Chicago Press, pp. 351–382. Anderson, C., and M. Wolff, 2010. “The Web Is Dead. Long Live the Internet,” Wired, September. http://www.wired.com/magazine/2010/08/ff_webrip/. Arora, A., and F. Bokhari, 2007. “Open versus Closed Firms and the Dynamics of Industry Evolution,” Journal of Industrial Economics, 55(3), 499–527. Augereau, A., S. Greenstein, and M. Rysman, 2006. “Coordination versus Differentiation in a Standards War: 56K Modems,” Rand Journal of Economics, 34 (4), 889–911. Baumol, W., and twenty six economists. 2006. Economists’ Statement on U.S. Broadband Policy (March 2006). AEIBrookings Joint Center Working Paper No. 06-06-01. Available at SSRN: http://ssrn.com/abstract=892009. Berners-Lee, T., and M. Fischetti, 1999. Weaving the Web, The Original Design and Ultimate Destiny of the World Wide Web, Harper Collins, New York. Besen, S., P. Milgrom, B. Mitchell, and P. Sringanesh, 2001. “Advances in Routing Technologies and Internet Peering Agreements,” American Economic Review, May, pp. 292–296. Blumenthal, M. S., and D. D. Clark, 2001. “Rethinking the Design of the Internet: The End-to-End Arguments vs. The Brave New World.” In (eds.) B. Compaine and S. Greenstein, Communications Policy in Transition: The Internet and Beyond. MIT Press: Cambridge, Mass., pp. 91–139. Bresnahan, T., 1999. “The Changing Structure of Innovation in Computing” in (ed.) J. A. Eisenach and T. M. Lenard, Competition, Convergence and the Microsoft Monopoly: Antitrust in the Digital Marketplace, Kluwer Academic Publishers, Boston, pp. 159–208. (p. 29) Bresnahan, T., and Pai-Ling Yin, 2007. “Standard Setting in Markets: The Browser Wars,” in (eds.) S. Greenstein and V. Stango, Standards and Public Policy, Cambridge University Press; Cambridge, UK. pp. 18–59. Bresnahan, T., and S. Greenstein, 1999. “Technological Competition and the Structure of the Computer Industry,” Page 15 of 24 Internet Infrastructure Journal of Industrial Economics, March, pp. 1–40. Bresnahan, T., S. Greenstein, and R. Henderson, 2011. “Schumpeterian Competition and Diseconomies of Scope: Illustrations from the Histories of IBM and Microsoft.” Forthcoming in (eds.) J. Lerner and S. Stern, The Rate and Direction of Inventive Activity, 50th Anniversary, National Bureau of Economic Research. http://www.nber.org/books/lern11-1, accessed March, 2012. Burgelman, R., 2007. “Intel Centrino in 2007: A New Platform Strategy for Growth.” Stanford University Graduate School of Business, SM-156. Chiao, B., J. Tirole, and J. Lerner, 2007. “The Rules of Standard Setting Organizations,” Rand Journal of Economics, 34(8), 905–930. Clark, D., W. Lehr, S. Bauer, P. Faratin, R. Sami, and J. Wroclawski, 2006. “Overlays and the Future of the Internet,” Communications and Strategies, 63, 3rd quarter, pp. 1–21. Crandall, R., 2005. “Broadband Communications,” in (eds.) M. Cave, S. Majumdar, and Vogelsang, Handbook of Telecommunications Economics, pp. 156–187. Amsterdam, The Netherlands: Elsevier. Cusumano, M., and D. Yoffie, 2000. Competing on Internet Time: Lessons from Netscape and Its Battle with Microsoft. New York: Free Press. Dalle, J.-M., P. A. David, R. A. Ghosh, and F. Wolak, 2004. “Free & Open Source Software Developers and ‘the Economy of Regard’: Participation and Code-Signing in the Modules of the Linu Kernel,” Working paper, SIEPR, Stanford University, Open Source Software Project http://siepr.stanford.edu/programs/OpenSoftware_David/NSFOSF_Publications.html Dedrick, J., and J. West, 2001. “Open Source Standardization: The Rise of Linu in the Network Era,” Knowledge, Technology and Policy, 14 (2), 88–112. Doms, M., and C. Forman, 2005. “Prices for Local Area Network Equipment,” Information Economics and Policy, 17(3), 365–388. Downes, T., and S. Greenstein, 2002. “Universal Access and Local Internet Markets in the U.S.,” Research Policy, 31, 1035–1052. Downes, T., and S. Greenstein, 2007. “Understanding Why Universal Service Obligations May Be Unnecessary: The Private Development of Local Internet Access Markets.” Journal of Urban Economics, 62, 2–26. Evans, D., A. Hagiu, and R. Schmalensee (2006). Invisible Engines: How Software Platforms Drive Innovation and Transform Industries, MIT Press; Cambridge, Mass. Federal Communications Commission, 2010a. National Broadband Plan, Connecting America, http://www.broadband.gov/. Federal Communications Commission, 2010b. “In the Matter of Preserving the Open Internet Broadband Industry Practices,” GN Docket No. 09–191, WC Docket No. 07–52, December 23, 2010. http://www.fcc.gov/. Forman, C., and A. Goldfarb, 2006. “Diffusion of Information and Communications Technology to Business,” in (ed.) T. Hendershott, Economics and Information Systems, Volume 1, Elsevier, pp. 1–43. Forman, C., A. Goldfarb, and S. Greenstein, 2005. “How did Location Affect Adoption of the Internet by Commercial Establishments? Urban Density versus Global Village,” Journal of Urban Economics, 58(3), 389–420. (p. 30) Fosfuri, A., M. Giarratana, and A. Luzzi, 2005. “Firm Assets and Investments in Open Source Software Products.” Druid Working Paper No. 05–10. Copenhagen Business School. Friedan, R., 2001. “The Potential for Scrutiny of Internet Peering Policies in Multilateral Forums,” in (eds.) B. Compaine and S. Greenstein, Communications Policy in Transition, The Internet and Beyond, MIT Press; Cambridge, Mass., 159–194. Page 16 of 24 Internet Infrastructure Gawer, A., 2009, Platforms, Innovation and Competition, Northampton, Mass.: Edward Elgar. Gawer, A., and M. Cusumano, 2002. Platform Leadership: How Intel, Microsoft and Cisco Drive Innovation. Boston, Mass.: Harvard Business School Press. Gawer, A., and R. Henderson, 2007. “Platform Owner Entry and Innovation in Complementary Markets: Evidence from Intel.” Journal of Economics and Management Strategy, Volume 16 (1), 1–34. Gilles, J., and R. Cailliau, 2000, How the Web Was Born, Oxford, UK: Oxford University Press. Goldfarb, A., 2004. “Concentration in Advertising-Supported Online Markets: An Empirical Approach.” Economics of Innovation and New Technology, 13(6), 581–594. Goldstein, F., 2005. The Great Telecom Meltdown. Boston: Artech House. Greenstein, S., 2006. “Wikipedia in the Spotlight.” Kellogg School of Management, case 5–306–507, http://www.kellogg.northwestern.edu/faculty/kellogg_case_collection.aspx. Greenstein, S., 2007a. “Economic Experiments and Neutrality in Internet Access Markets,” in (eds.) A. Jaffe, J. Lerner and S. Stern, Innovation, Policy and the Economy, Volume 8. Cambridge, Mass.: MIT Press, pp. 59–109. Greenstein, S., 2007b. “The Evolution of Market Structure for Internet Access in the United States,” in (eds.) W. Aspray and P. Ceruzzi, The Commercialization of the Internet and its Impact on American Business. Cambridge, Mass.: MIT Press, pp. 47–104. Greenstein, S., 2009a. “Open Platform Development and the Commercial Internet.” In (ed.) A. Gawer, Platforms, Innovation and Competition, Northampton, Mass.: Edward Elgar, pp. 219–250. Greenstein, S., 2009b. “Glimmers and Signs of Innovative Health in the Commercial Internet,” Journal of Telecommunication and High Technology Law, pp. 25–78. Greenstein, S., 2010. “The Emergence of the Internet: Collective Invention and Wild Ducks.” Industrial and Corporate Change. 19(5), 1521–1562. Greenstein, S., 2011. “Nurturing the Accumulation of Innovations: Lessons from the Internet,” in (eds.) Rebecca Henderson and Richard Newell, Accelerating Innovation in Energy: Insights from Multiple Sectors, University of Chicago Press, pp. 189–224. Greenstein, S., and R. McDevitt. 2009. “The Broadband Bonus: Accounting for Broadband Internet's Impact on U.S. GDP.” NBER Working paper 14758. http://www.nber.org/papers/w14758. Greenstein, S., and R. McDevitt, 2011. “Broadband Internet's Impact on Seven Countries,” in (ed.) Randy Weiss, ICT and Performance: Towards Comprehensive Measurement and Analysis. Madrid: Fundacion Telefonic, pp. 35– 52. Greenstein, S., and J. Prince, 2007. “The Diffusion of the Internet and the Geography of the Digital Divide,” in (eds.) R. Mansell, C. Avgerou, D. Quah, and R. Silverstone, Oxford Handbook on ICTs, Oxford University Press, pp. 168– 195. (p. 31) Haigh, T., 2007. “Building the Web's Missing Links: Portals and Search Engines” in (eds.) W. Aspray and P. Ceruzzi, The Internet and American Business, MIT Press. Kahin, B., and B. McConnell, 1997. “Towards a Public Metanetwork; Interconnection, Leveraging and Privatization of Government-Funded Networks in the United States,” in (eds.) E. Noam and A. Nishuilleabhain, Private Networks Public Objectives, Elsevier, Amsterdam, the Netherlands. pp. 307–321. Kahn, R., 1995. “The Role of Government in the Evolution of the Internet,” in (ed.) National Academy of Engineering, Revolution in the U.S. Information Infrastructure. Washington, D.C.: National Academy Press, 13–24. Kende, M., 2000. “The Digital Handshake: Connecting Internet Backbones,” Working Paper No. 32, Federal Page 17 of 24 Internet Infrastructure Communications Commission, Office of Planning and Policy, Washington, D.C. Kesan, J. P., and R. C. Shah, 2001. “Fool Us Once, Shame on You—Fool Us Twice, Shame on Us: What We Can Learn from the Privatizations of the Internet Backbone Network and the Domain Name System,” Washington University Law Quarterly, 79, 89–220. Laffont, J.-J., J. S. Marcus, P. Rey, and J. Tirole, 2001. “Internet Peering,” American Economic Review, Papers and Proceedings, 91,. 287–292. Laffont, J.-J., J. S. Marcus, P. Rey, and J. Tirole, 2003. “Internet Interconnection and the Off-net Pricing Principle,” Rand Journal of Economics, 34(2), 370–390. Leiner, B., V. Cerf, D. Clark, R. Kahn, L. Kleinrock, D. Lynch, J. Postel, L. Roberts, and S. Wolff, 2003. A Brief History of the Internet, Version 3.32, Last revised, 10 December, 2003. http://www.isoc.org/internet/history/brief.shtml, downloaded August, 2009. Lerner, J., and J. Tirole, 2002. “Some Simple Economics of Open Source Software,” Journal of Industrial Economics, 50 (2), 197–234. Lessig, L., 1999, Code and Other Laws of Cyber Space, New York: Basic Books. Mackie-Mason, J., and J. Netz, 2007., “Manipulating Interface Standards as an Anticompetitive Strategy,” in (eds.) Shane Greenstein and Victor Stango, Standards and Public Policy, Cambridge Press; Cambridge, Mass., pp. 231– 259. Marcus, J. S., 2008. “IP-Based NGNs and Interconnection: The Debate in Europe,” Communications & Strategies, No. 72, p. 17. Marcus, J. S., and D. Elixmann, 2008. “Regulatory Approaches to Next Generation Networks (NGNs): An International Comparison.” Communications and Strategies, 69 (1st quarter), 1–28. Mockus, A., R. T. Fielding, and J. D. Herbsleb, 2005. “Two Case Studies of Open Source Software Development: Apache and Mozilla,” in (ed.) J. Feller, Perspectives on Free and Open Source Software. Cambridge, MA: MIT Press; Cambridge, pp 163–210. Mowery, D. C., and T. S. Simcoe, 2002a. “The Origins and Evolution of the Internet,” in R. Nelson, B. Steil, and D. Victor (eds.), Technological Innovation and Economic Performance. Princeton, N.J.: Princeton University Press, 229–264. Mowery, D. C. and T. S. Simcoe, 2002b. “Is the Internet a U.S. Invention? An Economic and Technological History of Computer Networking.” Research Policy, 31(8–9), 1369–1387. Mueller, M.L., 2002. Ruling the Root: Internet Governance and the Taming of Cyberspace, Cambridge, Mass.: MIT Press. (p. 32) NTIA, 2010. “Exploring the Digital Nation: Home Broadband Internet Adoption in the United States,” http://www.ntia.doc.gov/reports.html. Nuechterlein, J. E., and P. J. Weiser, 2005. Digital Crossroads: American Telecommunications Policy in the Internet Age, Cambridge, Mass.: MIT Press. Odlyzko, A., 2010. “Bubbles, Gullibility, and Other Challenges for Economics, Psychology, Sociology, and Information Sciences.” First Monday 15, no. 9. http://www.uic.edu/htbin/cgiwrap/bin/ojs/index.php/fm/article/view/3142/2603 Ou, G., 2008. “A Policy Maker's Guide to Network Management. The Information Technology and Innovation Foundation,” http://www.itif.org/index.php?id=205. Oxman, J., 1999. The FCC and the Unregulation of the Internet. Working paper 31, Federal Communications Commission, Office of Planning and Policy, Washington, D.C. Page 18 of 24 Internet Infrastructure Partridge, C., 2008. “The Technical Development of Internet e-Mail,” IEEE Annals of the History of the Computing 30, pp. 3–29. Quarterman, J.S., 1989. Matri Computer Networks and Conferences, Bedford, Mass.: Digital Press. Rosston, G., 2009. The Rise and Fall of Third Party High Speed Access.” Information, Economics and Policy 21, pp. 21–33. Rosston, G., S. J. Savage, and D. Waldman, 2010. Household Demand for Broadband Internet Service.” SIEPR Paper 09–008. http://siepr.stanford.edu/publicationsprofile/2109. Russell, A.L., 2006. Rough Consensus and Running Code and the Internet-OSI Standards War.” IEEE Annals of the History of Computing, 28, pp. 48–61. Saltzer, J.H., D.P. Reed, and D.D. Clark, 1984. “End-to-End Arguments in System Design,” ACM Transactions on Computer Systems 2, pp. 277–288 . An earlier version appeared in the Second International Conference on Distributed Computing Systems (April, 1981) pp. 509–512. Savage, S. J., and D. Waldman, 2004, “United States Demand for Internet Access,” Review of Network Economics 3, pp. 228–247. Seamans, R., 2010. “Fighting City Hall: Entry Deterrence and New Technology Upgrades in Local Cable TV Markets.” Working paper, NYU Stern School of Business, http://pages.stern.nyu.edu/~rseamans/index.htm. Simcoe, T., 2007. “Delay and De jure Standardization: Exploring the Slow Down in Internet Standards Development,” in (eds.) Shane Greenstein and Victor Stango, Standards and Public Policy, Cambridge, Mass.: Cambridge Press, pp. 260–297. Simcoe, T., 2010. “Standard Setting Committees: Consensus Governance for Shared Technology Platforms,” Working Paper, Boston University, http://people.bu.edu/tsimcoe/index.html. Stranger, G., and S. Greenstein, 2007. “Pricing in the Shadow of Firm Turnover: ISPs in the 1990s.” International Journal of Industrial Organization, 26, pp. 625–642. Strover, S., 2001. “Rural Internet Connectivity.” Telecommunications Policy, 25, pp. 331–347. Von Burg, U., 2001. The Triumph of Ethernet: Technological Communities and the Battle for the LAN Standard, Palo Alto, CA: Stanford University Press. Von Hippel, E., 2005. “Open Source Software Projects as User Innovation Networks,” in (ed.) Joseph Feller, Brian Fitzgerald, Scott A. Hissam, Karim Lakhani, Perspectives on Free and Open Software, Cambridge, Mass.: MIT Press; pp. 267–278. Von Schewick, B., 2010. Internet Architecture and Innovation, Cambridge, Mass.: MIT Press. (p. 33) Waldrop, M., 2001. The Dream Machine: J.C.R. Licklider and the Revolution that Made Computing Personal, New York: Penguin. Wallsten, S., 2009. “Understanding International Broadband Comparison,” Technology Policy Institute, http://www.techpolicyinstitute.org/publications/topic/2.html West, J., and S. Gallagher, 2006. Patterns of Innovation in Open Source Software, in (ed.) Hank Chesbrough, Wim Haverbeke, Joel West, Open Innovation: Researching a New Paradigm, Oxford University Press: Oxford, pp. 82– 108. Zittrain, J., 2009. “The Future of the Internet and How to Stop It, Creative Commons Attribution—Noncommercial Share Alike 3.0 license,” www.jz.org. Page 19 of 24 Internet Infrastructure Notes: (1.) This is explained in considerable detail in Ou (2008). (2.) There are many fine books about these developments, including Abbate (1999), Leiner et al. (2003), and Waldrop (2001). (3.) Packet switching had been discussed among communications theorists since the early 1960s, as well as by some commercial firms who tried to implement simple versions in their frontier systems. As has been noted by others, the ideas behind packet switching had many fathers: Paul Baran, J. C. Likelider, Douglas Engelbart, and Len Kleinrock. There are many accounts of this. See, e.g., Quarterman (1989), Abbate (1999), Leiner et al. (2003), or Waldrop (2001). (4.) These advances have been documented and analyzed by many writers, including, e.g., Quarterman (1989), Abbate (1999), Leiner et al. (2003), and Waldrop (2001). (5.) An extensive explanation of TCP/IP can be found in many publications. Summarized simply, TCP determined a set of procedures for moving data across a network and what to do when problems arose. If there were errors or specific congestion issues, TCP contained procedures for retransmitting the data. While serving the same function of a postal envelope and address, IP also shaped the format of the message inside. It specified the address for the packet, its origin and destination, a few details about how the message format worked, and, in conjunction with routers, the likely path for the packet toward its destination. See, e.g., Leiner et al. (2003). (6.) In part, this was due to a DOD requirement that all Unix systems do so, but it also arose, in part, because most Unix users in research environments found this feature valuable. (7.) Moreover, a set of fail-safes had been put in place to make sure that one error message did not trigger another. That is, the system avoided the nightmare of one error message generating yet another error message, which generated another and then another, thus flooding the system with never-ending messages. (8.) The paper that defined this phrase is commonly cited as Saltzer et al. (1984). (9.) As stated by Blumenthal and Clark (2001) in a retrospective look: “When a general-purpose system (for example, a network or an operating system) is built and specific applications are then built using this system (for example, email or the World Wide Web over the Internet), there is a question of how these specific applications and their required supporting services should be designed. The end-to-end arguments suggest that specific application-level functions usually cannot, and preferably should not, be built into the lower levels of the system— the core of the network.” (10.) End-to-end also differed from most large computing systems at the time, such as mainframes, which put the essential operations in the central processing unit. When such computers became situated in a networked environment, there was little else for the terminals to do. They became known as dumb terminals. Contrast with end-to-end, which colloquially speaking, located intelligence at the edges of the system, namely, in the clients. (11.) The economics behind this “surprise” is discussed in some detail in Greenstein (2011). (12.) See, e.g., Aizcorbe et al. (2007), Doms and Forman (2005), or the discussion in Forman and Goldfarb (2006). (13.) See the discussions in, e.g., Kahn (1995), Kahin and McConnel (1997), and Kesan and Shah (2001), or Greenstein (2010). (14.) On many of the challenges during the transition, see Abbate (1999) or Kesan and Shah (2001). (15.) For longer discussions about the origins and economic consequences, see, e.g., Mowery and Simcoe (2002a, b), and Greenstein (2011). (16.) The term “mesh” is due to Besen et al. (2001). (17.) See Friedan (2001). Page 20 of 24 Internet Infrastructure (18.) For insights into the incentives to conduct traffic and come to peering agreements, see Besen et al. (2001) and Laffont et al. (2001, 2003). (19.) For more on overlays, see Clark et al. (2006). (20.) See the evidence in Rosston (2009). (21.) This topic has received attention ever since the commercial Internet first began to blossom. See, e.g., Strover (2001) or Downes and Greenstein (2002). A summary can be found in Greenstein and Prince (2007). Also see the discussion in, e.g., the Federal Communications Commission (2010a). (22.) See, e.g., Berners-Lee and Fischetti (1999), and Gilles and Cailliau (2000). (23.) For more on this story, see Cusamano and Yoffie (2000). (24.) Andrew Odlyzko maintains an accessible summary of studies and forecasts of data traffic at http://www.dtc.umn.edu/mints/. (25.) Anderson and Wolff (2010) present a very accessible version of this argument. (26.) Longer explanations for these events can be found in Greenstein (2007a, b). (27.) For a longer discussion, see, e.g., Downes and Greenstein (2002, 2007). (28.) For a discussion, see, e.g., Goldfarb (2004), Haigh (2007). (29.) High usage could happen for a variety of reasons. For example, some technical users simply enjoyed being online for large lengths of time, surfing the growing Internet and Web. Some users began to operate businesses from their homes, remaining online throughout the entire workday. Some users simply forgot to log off, leaving their computers running and tying up the telephone line supporting the connection to the PC. And some users grew more experienced, and found a vast array of activities more attractive over time. (30.) One survey of pricing contracts in May 1996 found that nearly 75 percent of the ISPs offering 28K service (the maximum dial-up speed at the time) offered a limited plan in addition to their unlimited plan. That dropped to nearly 50 percent by August. By March of 1997 it was 33 percent, 25 percent by January of 1998, and less than 15 percent by January of 1999. For a summary see Stranger and Greenstein (2007). (31.) During the 1990s most cable companies sold access to the line directly to users but made arrangements with other firms, such as Roadrunner or @home, to handle traffic, routing, management, and other facets of the user experience. Some of these arrangements changed after 2001, either due to managerial preferences, as when @home lost its contract, or due to regulatory mandates to give users a choice over another ISP, as occurred after the AOL/Time Warner merger. See Rosston (2009). (32.) Download speed may not reach the advertised maxima. In cable networks, for example, congestion issues were possible during peak hours. In DSL networks, the quality of service could decline significantly for users far away from the central switch. The results are difficult to measure with precision. (33.) The descriptive results were published in reports authored by staff at the NTIA. See NTIA (2010). (34.) In addition to the surveys by Pew, also see, e.g., Savage and Waldman (2004), Rosston et al. (2010), and the summary of Greenstein and Prince (2007). For surveys of business adoption and its variance over geography, see Forman and Goldfarb (2006) and Forman et al. (2005). (35.) See, e.g., Crandall (2005). (36.) In many areas, fiber to the home was prohibitively expensive for almost all users except businesses, and even then, it was mostly used by businesses in dense urban areas, where the fiber was cheaper to lay. Fiber to the home has recently become cheaper and may become a viable option sometime in the future. See Crandall (2005). (37.) OECD Broadband Portal, http://www.oecd.org/sti/ict/broadband, Table 1d. Page 21 of 24 Internet Infrastructure (38.) OECD Broadband Portal, http://www.oecd.org/sti/ict/broadband, Table 1k. For a critique of the US standings in these rankings and insights into how to correct misunderstandings, see Wallsten (2009). Perhaps the biggest issue is the denominator, which is per capita. However, a subscriber tends to subscribe to one line per household. Since US average household size is much larger than average household size in other countries, the figure gives the false impression that US residences have less access to broadband than is actually accurate. (39.) These come from Greenstein and McDevitt (2011). (40.) For sources, see Greenstein and McDevitt (2009, 2011). (41.) See e.g., Federal Communications Commission (2010b), or Marcus (2008). (42.) See e.g., http://www.wispa.org/. Also see http://wispassoc.com/. (43.) For interpretations of platform incentives, see e.g., Gawer (2009) or Evans, Haigu and Schmalensee (2006). (44.) This is distinct from an engineering notion of a platform. The designers of the Internet deliberately built what they regarded as a computing platform. The inventors employed what they regarded as a balanced and standardized bundle of components to regularly deliver services. This balance reflected a long-standing and familiar principle in computer science. The inventors and DARPA administrators anticipated a benefit from this design: others would build applications, though these inventors did not presume to know what those applications would do specifically. (45.) See, e.g., Bresnahan and Greenstein (1999), or Dedrick and West (2001). (46.) See, e.g., Bresnahan (1999). (47.) See, e.g., Gawer and Cusumano (2002) or Gawer and Henderson (2007). (48.) Sometimes Google retains many proprietary features, particularly in its search engine, which also supports a lucrative ad-placement business. Google takes action to prevent anyone from imitating it. For example, the caching, indexing, and underlying engineering tweaking activities remain hidden from public view. (49.) See e.g., Cusumano and Yoffie (2000), Bresnahan and Yin (2007), or Bresnahan et al. (2011). (50.) For an account of this decision, see Burgelman (2007). (51.) For a book with this theme, see, e.g., Zittrain (2009), Lessig (1999), or Von Schewick (2010). (52.) Anderson and Wolff (2010) or Zittrain (2009). (53.) Once again, this case is explained in detail in Greenstein (2009b). (54.) Numerous computer scientists and networking experts have pointed out that both Sprint and Cogent could have adjusted their routing tables in advance to prevent users from being cutoff. Hence, there is a sense in which both parties bear responsibility for imposing costs on their users. (55.) It appears that Sprint's capitulation to its user base is, however, evidence that Sprint's management does not have the ability to ignore its users for very long. (56.) This case is explained in detail in Greenstein (2009b). (57.) For summary, see Adam Rothschild, December 2, 2010, http://www.voxel.net/blog/2010/12/peering-disputescomcast-level-3-and-you. (58.) This is especially true of the Internet Architecture Board and IETF, before it moved under the auspices of the Internet Society in 1992, where it remains today. See, e.g., Abbate (1999) or Russell (2006). (59.) See Abbate (1999) for a history of the design of these protocols. See Partridge (2008) for a history of the processes that led to the development of email, for example. Page 22 of 24 Internet Infrastructure (60.) Simcoe (2007, 2010) provides an overview of the operations at IETF and its changes as it grew. (61.) For further discussion see Farrell and Simcoe, chapter 2 of this volume. (62.) For example, see, e.g., Chiao et al. (2007). (63.) These contrasts are further discussed in Greenstein (2009a). (64.) See, e.g., Russell (2006). (65.) The latter forbearance was deliberate. On the lack of interference in the design of the Ethernet, see von Burg (2001). On the design of 56K modems, see Augereau, Greenstein, and Rysman (2007). On the lack of regulation for network interconnection, see the full discussions in, e.g., Oxman (1999) or Kende (2000) or the account in Neuchterlein and Weiser (2005). More recent experience has departed from these trends, particularly in principles for regulating last-mile infrastructure. A summary of these departures is in Greenstein (2007b). (66.) For a longer explanation of these origins, see, e.g., Kesan and Shah (2001) or Mueller (2002). (67.) There is considerable writing about the growth of the production of Linux software and from a variety of perspectives. See, e.g., Dalle, David, Ghosh, and Wolak (2004), Lerner and Tirole (2002), VonHippel (2005), West and Gallagher (2006), or Arora and Farasat (2007). For an account and analysis of how many firms got on the Linux bandwagon, see, e.g., Dedrick and West (2001) or Fosfuri, Giarratana and Luzzi (2005). For further discussion see chapter 17 by Johnson, this volume. (68.) For more on the history and operation of Apache, see, e.g., Mockus, Fielding, and Herbsleb (2005). (69.) For more information, see Greenstein (2006). (70.) For further discussion, see the chapter 3 by Hagiu, this volume. (71.) The figures come from the website maintained by the Creative Commons, http://creativecommons.org/. (72.) For one interesting account of the changing ratio of “suits to beards” at the IETF, see Simcoe (2007, 2010). For an account of the manipulation of hearings at the IEEE, see Mackie-Mason and Netz (2007). (73.) Federal Communications Commission (2010a). (74.) See Marcus (2008) and Marcus and Elixmann (2008). (75.) The overinvestment in Internet infrastructure in the late 1990s had many causes. These are analyzed by, among others, Goldstein (2005), Greenstein (2007b), and Odlyzko (2010). (76.) Statistics in the National Broadband Plan, FCC (2010a), seem to indicate that 3 percent of US homes subscribe to this service as of the end of 2009. (77.) The economics behind the high cost of providing broadband in low-density locations is explained in detail in Strover (2001), Crandall (2005), and Greenstein and Prince (2007). (78.) See Marcus (2008) and Marcus and Elixmann (2008). (79.) For a summary, see Goldstein (2005), Nuechterlein and Weiser (2005), Greenstein (2010), and Federal Communications Commission (2010b). (80.) See, e.g., Seamans (2010). (81.) See, e.g., Baumol et al. (2006). Shane Greenstein Prof Shane Greenstein is Kellogg Chair in Information Technology at the Kellogg School of Management, Northwestern University, Evanston, USA Page 23 of 24 Four Paths to Compatibility Oxford Handbooks Online Four Paths to Compatibility Joseph Farrell and Timothy Simcoe The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0002 Abstract and Keywords This article describes the economics of standardization, and the costs and benefits of alternative ways to achieve compatibility. Four paths to compatibility, namely standards wars, negotiations, dictators, and converters, are explained. These four paths to compatibility have different costs and benefits. Standard setting organizations are a heterogeneous set of institutions connected by their use of the consensus process. Government involvement may be appropriate when private control of an interface would result in utmost market power. Converters are attractive because they preserve flexibility for implementers. Compatibility standards can emerge through market competition, negotiated consensus, converters, or the actions of a dominant firm. Developing a better understanding of how a particular path is selected shows a crucial first step toward measuring the cost-benefit tradeoffs across paths, and adjudicating debates over the efficiency of the selection process. Keywords: standards wars, negotiations, dictators, converters, compatibility standards, standardization, standard setting organizations 1. Introduction Compatibility standards are design rules that promote product interoperability, such as the thread size for mechanical nuts and bolts or the communication protocols shared by all Internet devices. Products that adhere to standards should work together well, which produces a range of benefits: users may share information, or “mix and match” components; the cost of market entry declines; and there is a division of labor, thus enabling specialization in component production and innovation. This chapter describes four paths to compatibility— standards wars, negotiations, dictators, and converters—and explores how and when they are used, as alternatives or in combination. While product interoperability may pose engineering challenges, we focus on issues of economic incentive that arise when its costs and benefits are not evenly distributed. For example, firms that control a technology platform may resist compatibility with other systems or standards that could reduce switching costs for their installed base. Specialized component producers may fight against standards that threaten to “commoditize” their products. Even when support for compatibility is widespread, rival firms may advocate competing designs that confer private benefits because of intellectual property rights, lead-time advantages, or proprietary complements. Given this mix of common and conflicting interests, we focus on four natural ways to coordinate design decisions (in the sense of achieving compatibility). The first (p. 35) is decentralized choice, which can yield coordinated outcomes when network effects are strong, even if the resulting process is messy. Negotiations are a second coordination mechanism. In particular, firms often participate in voluntary standard setting organizations (SSOs), which seek a broad consensus on aspects of product design before endorsing a particular technology. A third route to compatibility is to follow the lead of an influential decision maker, such as a large customer or platform leader. Finally, participants may abandon efforts to coordinate on a single standard and instead patch together 1 Page 1 of 20 Four Paths to Compatibility partial compatibility through converters or multihoming.1 These four paths to compatibility have different costs and benefits, which can be measured in time and resources, the likelihood of successful coordination for compatibility, and the ex post impact on competition and innovation. Whether these complex welfare trade-offs are well internalized depends on how (and by whom) the path to compatibility is chosen. A full treatment of the compatibility problem would specify the selection process and quantify the relative performance of each path. In practice, although theory clarifies the potential trade-offs, we have limited empirical evidence on the comparative costs and benefits of each path, or the precise nature of the selection process. Sometimes the choice of a particular path to compatibility is a more-or-less conscious decision. For example, firms can decide whether to join the deliberations of an SSO or follow the lead of a dominant player. A dominant player can decide whether to commit to a standard and expect (perhaps hope) to be followed, or defer to a consensus negotiation. As these examples suggest, it can be a complex question who, if anyone, “chooses” the mechanism, if any, used to coordinate. Some market forces push toward efficiency, but it is not guaranteed. For example, a platform leader has a general incentive to dictate efficient interface standards or to allow an efficient evolution process, but that incentive may coexist with, and perhaps be overwhelmed by, incentives to stifle ex post competition. Likewise, competition among SSOs may or may not lead them toward better policies, and standards wars may or may not tip toward the superior platform. Sometimes firms will start down one path to compatibility and then veer onto another. For instance, a decentralized standards war may be resolved by resort to an SSO or through the intervention of a dominant firm. Slow negotiations within an SSO can be accelerated by evidence that the market is tipping, and platform sponsors may promote complementary innovation by using an SSO to open parts of their platform. Although theory suggests that certain “hybrid paths” can work well, we know rather little about how different coordination mechanisms complement or interfere with one another. This chapter begins by explaining something familiar to many readers: how the choice of interoperability standards resembles a coordination game in which players have a mix of common and conflicting incentives. In particular, it explains how compatibility produces broadly shared benefits, and discusses several reasons that firms may receive private benefits from coordinating on a preferred technology. Section 3 describes costs and benefits of our four paths to compatibility. Section 4 examines the selection process and the role of “hybrid” paths to compatibility. Section 5 concludes. (p. 36) 2. Costs and Benefits of Compatibility When all influential players favor compatibility, creating or upgrading standards involves a coordination problem. When there is but one technology, or when participants share common goals and notions of quality, the solution is primarily a matter of communication that can be solved by holding a meeting, or appointing a focal adopter whom all agree to follow. But if there are several technologies to choose from and participants disagree about their relative merit, it turns a pure coordination game into a battle of the sexes, where players may try to “win” by arguing for, or committing to, their preferred outcome. Figure 2.1 illustrates the basic dilemma in a symmetric two-player game. As long as C 〉 B, the benefits of compatibility outweigh the payoffs from uncoordinated adoption of each player's preferred technology, and the game has two pure-strategy Nash Equilibria: both adopt A, and both adopt B. Each player gains some additional private benefit (equal to D) in the equilibrium that selects their preferred technology. When these private benefits are small (D ≈ 0), many coordination mechanisms would work well. But as D grows large, players will push hard for their preferred equilibrium. Whether these efforts to promote a particular outcome are socially productive depends on a variety of factors about the players’ available actions and the details of the equilibrium selection process. Later, we assume that D 〉 0, and compare the costs and benefits of four broad methods for choosing an equilibrium. But first, this section explains why the payoffs in Figure 2.1 can be a sensible way to model the choice of compatibility standards, particularly in the information and communications technology (ICT) sector. Page 2 of 20 Four Paths to Compatibility Click to view larger Figure 2.1 Compatibility Choice as a Coordination Game. The benefits of compatibility (C-B in Figure 2.1) come in two flavors: horizontal and vertical. Horizontal compatibility is the ability to share complements across multiple platforms, and we call a platform horizontally open if its installed base of complements can be easily accessed from rival systems. Many parts of the Internet are in this sense horizontally open. For example, web pages can be displayed on (p. 37) competing web browsers, and rival instant messenger programs allow users to chat. Video game consoles and proprietary operating systems, such as Microsoft Windows, by contrast, are horizontally closed: absent further action (such as “porting”), an application written for one is not usable on others. These distinctions can be nuanced. For example, what if a platform's set of complements is readily available to users of rival platforms, but at an additional charge, as with many banks’ ATM networks? Similarly, Microsoft may have choices (such as the degree of support offered to cross-platform tools like Java) that affect, but do not fully determine, the speed and extent to which complements for Windows become available on other platforms. Benefits of horizontal compatibility include the ability to communicate with a larger installed base (direct network effects) and positive feedback between the size of an installed base and the supply of complementary goods (indirect network effects). Katz and Shapiro (1985) analyzed oligopoly with firm-specific demand-side increasing returns. A more recent literature on many-sided platforms (e.g. Rochet and Tirole 2003; Parker and Van Alstyne 2005; Weyl 2010) extends the analysis of indirect network effects by allowing externalities and access prices to vary across different user groups.2 Vertical compatibility is the ability of those other than the platform sponsor to supply complements for the system. We call a platform vertically open if independent firms can supply complements without obtaining a platform leader's permission.3 For example, the Hush-a-Phone case (238 F.2d 266, 1956), and the FCC's later Carterfone decision (13 F.C.C.2d 420) opened the U.S. telephone network to independently supplied attachments such as faxes, modems, and answering machines. Many computing platforms, including Microsoft Windows, use vertical openness to attract independent software developers. Like horizontal openness, vertical openness can be a matter of degree rather than a sharp distinction. For instance, a platform leader may offer technically liberal access policies but charge access fees. Vertical compatibility produces several types of benefits. There are benefits from increased variety when vertical compatibility allows users to “mix and match” components (Matutes and Regibeau 1988). Vertical openness also can reduce the cost of entry, strengthening competition in complementary markets. Finally, vertical compatibility leads to a “modular” system architecture and division of innovative labor. Isolating a module that is likely to experience a sustained trajectory of improvements allows other components to take advantage of performance gains while protecting them from the cost of redesign. And when the locus of demand or the value of complementary innovations is highly uncertain, modularity and vertical openness facilitate simultaneous design experiments (Bresnahan and Greenstein 1999; Baldwin and Clark 2000). The benefits of horizontal or vertical compatibility are often broadly shared but need not be symmetric: there can also be private benefits of having a preferred technology become the industry standard. Such private benefits, labeled D in Figure 2.1, often lead to conflict and coordination difficulties in the search for compatibility. (p. 38) One important source of conflict is the presence of an installed base. Upgrading an installed base can be costly, and firms typically favor standards that preserve their investments in existing designs. Moreover, platform leaders with a large installed base will favor designs that preserve or increase switching costs, while prospective Page 3 of 20 Four Paths to Compatibility entrants push to reduce them. For example, in its US antitrust case, Microsoft was convicted of using illegal tactics to prevent Windows users, developers, and OEMs from migrating to the independent standards embodied in the Netscape browser and Java programming language. Design leads are another source of conflict. Short ICT product life cycles leave firms a limited window of opportunity to capitalize on the demand unleashed by a new standard, and first-mover advantages can be important. Thus, firms may try to block or delay a new standard if rivals have a significant lead at implementation. DeLacy et al. (2006) describe such efforts in the context of Wi-Fi standards development. In some cases, there is conflict over the location of module boundaries, or what engineers call the “protocol stack.” Since compatibility often promotes entry and competition, firms typically prefer to standardize components that complement their proprietary technology, but leave room for differentiation in areas where they have a technical edge. For example, Henderson (2003) describes how the networking technology start-up Ember allegedly joined several SSOs to prevent new standards from impinging on its core technology. Conflicts can also emerge when firms own intellectual property rights in a proposed standard, which they hope to license to implementers or use as leverage in future negotiations. Lerner, Tirole, and Strojwas (2003) show that nearly all “modern” patent pools are linked to compatibility standards, and Simcoe (2007) documents a rapid increase in intellectual property disclosures in the formal standard-setting process. While data on licensing are scant, Simcoe, Graham, and Feldman (2009) show that patents disclosed in the formal standards process have an unusually high litigation rate. Finally, conflicting interests can amplify technological uncertainty. In particular, when technical performance is hard to measure, firms and users will grow more skeptical of statements from self-interested participants about the quality of their favored design. 3. Four Paths to Compatibility Given this mix of conflict, common interest, and incomplete information, choosing compatibility standards can be a messy process. This section considers the performance of four paths to compatibility—standards wars, SSOs, dictators, and converters—in terms of the probability of achieving compatibility, the expected time and resource costs, and the implications for ex post competition and innovation. (p. 39) We find that economic theory helps articulate some of the complex trade-offs among these paths, but there is little systematic evidence. 3.1. Standards Wars Standards wars can be sponsored or unsponsored; for brevity we focus here on the sponsored variety, in which proponents of alternative technologies seek to preempt one another in the marketplace, each hoping that decentralized adoption will lead to their own solution becoming a de facto standard through positive feedback and increasing returns. Standards wars have broken out over video formats, modem protocols, Internet browsers, and transmission standards for electricity and cellular radio. These wars can be intense when horizontally incompatible platforms compete for a market with strong network effects, which they expect to tip toward a single winner who will likely acquire market power. Much has been written about the tactics and outcomes in such wars, but we do not attempt to discuss them comprehensively here, only to remind the reader of some of the dynamics.4 Standards wars often involve a race to acquire early adopters and efforts to manipulate user expectations, as described in Besen and Farrell (1994) or Shapiro and Varian (1998). Preemption is one strategy for building an early lead in an adoption race. Another strategy is to aggressively court early adopters with marketing, promotions, and pricing. Firms may also work to influence users’ expectations regarding the likely winner, since these beliefs may be self-fulfilling.5 Firms that fall behind in a race for early adopters or expectations may use backward compatibility or bundling to catch up. Backward compatibility jump-starts the supply of complements for a new platform. For instance, many video game platforms can play games written for older consoles sold by the same firm. Bundling promotes the adoption of new standards by linking them to existing technology upgrades. For example, Sony bundled a Blu-ray disc player with the Playstation game console to promote that video format over HD-DVD, and Bresnahan and Yin Page 4 of 20 Four Paths to Compatibility (2007) argue that Microsoft took advantage of the Windows upgrade cycle to overtake Netscape in the browser wars. Given the range of tactics used in a standards war, does decentralized technology adoption provide an attractive route to coordination? One social cost is that it will often lead to the emergence of a horizontally closed platform. While one might question the direction of causality (perhaps intractable conflicts over horizontal interoperability lead to standards wars), alternative paths to coordination may produce more ex post competition and reduce the risk of stranded investments. The economic logic of standards wars seems consistent with concerns that markets may “tip” prematurely toward an inferior solution. While many cite the QWERTY keyboard layout (the standard in English-language keyboards) as an example (e.g. David 1990), Liebowitz and Margolis (1990) dispute the empirical evidence and suggest that markets will typically coordinate on the best available technology, as long as the benefits of changing platforms outweigh any switching (p. 40) costs. It is difficult to find natural experiments that might resolve this debate, for example by randomly assigning an early lead in settings where there are clear differences in platform quality. But even if standards wars typically “get it right” in terms of selecting for quality, coordination problems may affect the timing of standards adoption.6 Optimists argue that standards wars are speedy, since participants have strong incentives to race for early adopters, and that fierce ex ante competition offsets any social cost of ex post incompatibility. But competition for early adopters does not always take the form of creating and transferring surplus, and its benefits must be weighed against the costs of stranded investments in a losing platform. Moreover, the uncertainty created by a standards war may cause forward-looking users, who fear choosing the losing platform, to delay commitments until the battle is resolved. For example, Dranove and Gandal (2003) find that preannouncement of the DivX format temporarily slowed the adoption of the digital video disc (DVD). Augereau, Rysman, and Greenstein (2006) suggest that the standards war in 56K modems also delayed consumer adoption. When it is costly to fight a standards war, participants may seek an escape route, such as some type of truce.7 For example, the 56K modem standards war ended in adoption of a compromise protocol incorporating elements of both technologies. The battle between code-division multiple access (CDMA) and time division multiple access (TDMA) cellular phone technology ended in a duopoly stalemate, with each standard capturing a significant share of the global market. Ironically, these escape routes and stalemates illustrate a final strength of decentralized adoption as a path to compatibility: it can reveal that network effects are weak, or that technologies initially perceived as competing standards can ultimately coexist by serving different applications. For example, Bluetooth (IEEE 802.15) was conceived as a home-networking standard, but ceded that market to Wi-Fi (IEEE 802.11) and is now widely used in short-range low-power devices, such as wireless headsets, keyboards, and remote controls. Similarly, the plethora of digital image formats (JPEG, GIF, TIFF, PNG, BMP, etc.) reflect trade-offs between image quality and compression, as well as compatibility with specific devices. Since “war” is a poor metaphor for the process of matching differentiated technology to niche markets, Updegrove (2007) has proposed the alternative label of “standards swarms” for settings where network effects are weak relative to the demand for variety. 3.2. Standard Setting Organizations One alternative to standards wars is for interested parties to try and coordinate through negotiation. This process is often called formal or de jure standard setting, and typically occurs within consensus standard setting organizations. There are hundreds of SSOs, and many of these nonprofit institutions develop standards for safety and performance measurement, as well as product (p. 41) compatibility.8 We use a broad definition of SSO that includes globally recognized “big I” standard setters, such as International Telecommunication Union (ITU) and (International Organization for Standardization (ISO); private consortia that manage a particular platform, such as the Internet Engineering Task Force (IETF) and World Wide Web Consortium (W3C); and smaller consortia that focus on a particular technology, such as the USB Forum or the Blu-ray Disc Association.9 This definition could even be stretched to include collaborative product-development groups, such as open-source software Page 5 of 20 Four Paths to Compatibility communities. While the largest SSOs have hundreds of subcommittees and maintain thousands of specifications, small consortia can resemble joint ventures, wherein a select group of firms develop and cross-license a single protocol under a so-called promoter-adopter agreement. Standards practitioners typically distinguish between consortia and “accredited” standards developing organizations (SDOs). SDOs sometimes receive preferential treatment in trade, government purchasing, and perhaps antitrust in return for adhering to best practices established by a national standards agency, such as the American National Standards Institute (ANSI).10 Table 2.1 hints at the size and scope of formal standard-setting in the United States by counting entries in the 2006 ANSI catalog of American National Standards and listing the 20 largest ANSI-accredited SDOs.11 SSOs use a consensus process to reach decisions. Though definitions vary, consensus typically implies support from a substantial majority of participants. For example, most accredited SDOs require a super-majority vote and a formal response to any “good faith” objections before approving a new standard. Since SSOs typically lack enforcement power, this screening process may serve as a signal of members’ intentions to adopt a standard, or an effort to sway the market's beliefs. Rysman and Simcoe (2008) provide some empirical evidence that SSOs’ nonbinding endorsements can promote technology diffusion by studying citation rates for US patents disclosed in the standard-setting process, and showing that an SSO endorsement leads to a measurable increase in forward citations. Beyond using a loosely defined consensus process and relying on persuasion and network effects to enforce their standards, SSOs’ internal rules and organization vary widely. Some are open to any interested participant, while others charge high fees and limit membership to a select group of firms. Some SSOs have a completely transparent process, whereas others reveal little information. Some SSOs require members to grant a royalty-free license to any intellectual property contained in a standard, whereas others are closely aligned with royalty-bearing patent pools. There has been little empirical research on the internal organization of SSOs, but Lemley (2002) and Chiao, Lerner, and Tirole (2007) examine variation in SSOs’ intellectual property rights policies. Given SSOs’ heterogeneity, what can we say about the costs and benefits of the consensus process as a path to coordination? Since SSOs encourage explicit comparisons and often have an engineering culture that emphasizes the role of technical quality, there is some reason to expect higher-quality standards than would emerge from a standards war or an uninformed choice among competing (p. 42) technologies. This prediction appears in the stochastic bargaining model of Simcoe (2012), as well as the war-of-attrition model of Farrell and Simcoe (2012), where SSOs provide a quality-screening mechanism. Page 6 of 20 Four Paths to Compatibility Table 2.1 Major ANSI Accredited SSOs Acronym Standards ICT Full Name INCITS 10,503 Y International Committee for Information Technology Standards ASTM 8,339 N American Society for Testing and Materials IEEE 7,873 Y Institute of Electrical and Electronics Engineers UL 7,469 N Underwriters Laboratories ASME 7,026 N American Society of Mechanical Engineers ANSI/TIA 4,760 Y Telecommunications Industry Association ANSI/T1 3,876 Y ANSI Telecommunications Subcommittee ANSI/ASHRAE 3,070 N American Society of Heating, Refrigerating and Air-Conditioning Engineers AWS 2,517 N American Welding Society ANSI/NFPA 2,365 N National Fire Protection Association ANSI/EIA 2,011 Y Electronic Industries Association ANSI/SCTE 1,803 Y Society of Cable Telecommunications Engineers ANSI/AWWA 1,759 N American Water Works Association ANSI/AAMI 1,621 Y American Association of Medical Imaging ANSI/NSF 1,612 N National Sanitation Foundation ANSI/ANS 1,225 N American Nuclear Society ANSI/API 1,225 N American Petroleum Institute ANSI/X9 940 N Financial Industry Standards ANSI/IPC 891 Y Association Connecting Electronics Industries ANSI/ISA 872 Y International Society of Automation Total ICT 30,786 43% Notes: List of largest ANSI accredited Standards Developing Organizations based on a count of documents listed in the 2006 ANSI catalog of American National Standards. The “Standards” column shows the actual document count. The “ICT” column indicates the authors' judgment as to whether that organization's primary focus is creating compatibility standards. Page 7 of 20 Four Paths to Compatibility But technical evaluation and screening for quality can impose lengthy delays, especially when the consensus process gives participants the power to block proposed solutions. A survey by the National Research Council (1990) found that standards practitioners viewed delays as a major problem, and Cargill (2001) suggests that the opportunity costs of delayed standardization explain a broad shift from accredited SDOs toward less formal consortia. Farrell and Saloner (1988) develop a formal model to compare outcomes in a standards war (grab-thedollar game) to an SSO (war of attrition). Their theory predicts a basic trade-off: the formal consensus (p. 43) process leads to coordination more often, while the standards war selects a winner more quickly. SSOs have sought ways to limit deadlocks and lengthy delays. First, some grant a particular party the power to break deadlocks, though such unilateral decisions could be viewed as a distinct route to compatibility (see below). A second approach is to start early in the life of a technology, before firms commit to alternative designs. Illustrating the impact of commitment on delays, Simcoe (2012) shows how delays at the IETF increased as the Internet matured into a commercial platform. But early standardization also has downsides; in particular, private actors have little incentive to contribute technology if they see no commercial opportunity, so anticipatory standards rely heavily on participation from public sector institutions such as academia or government labs. A third way to resolve deadlocks is to agree on partial or incomplete standards. Such standards often include “vendor-specific options” to facilitate product differentiation. And SSO participants sometimes agree to a “framework” that does not achieve full compatibility but standardizes those parts of an interface where compromise can be reached (thus lowering the ex post cost of achieving compatibility through converters).12 Fourth, SSOs may work faster if competing interests are placed in separate forums, such as independent working groups within a large SSO or even independent consortia. Lerner and Tirole (2006) model forum shopping when there is free entry into certification, and show that technology sponsors will choose the friendliest possible SSO, subject to the constraint that certification sways user beliefs enough to induce adoption. This “competing forums” approach works well if there is demand for variety and converters are cheap. But if network effects are strong, forum shopping may produce escalating commitments in advance of a standards war. For example, the Blu-ray and HD-DVD camps each established an independent implementers’ forum to promote their own video format. Beyond providing a forum for negotiation and certification activities, SSOs are often a locus of collaborative research and development. Thus, one might ask whether selecting this path to coordination has significant implications for innovation. Some forms of innovation within SSOs raise a public goods problem: incentives are weak if all firms have free access to improvements, especially in highly competitive industries. Weiss and Toyofuku (1996) gather evidence of free riding in 10BaseT standards development. Cabral and Salant (2008) study a model where standardization leads to free riding in research and development (R&D), and find that firms may favor incompatibility if it helps them sustain a high rate of innovation. Eisenmann (2008) suggests that SSOs often struggle with “architectural” innovations that span many component technologies, since it is difficult to coordinate the decisions of specialized firms with narrow interests in the outcomes of a particular working group or technical committee. However, such problems need not prevent all innovation within SSOs. Firms often contribute proprietary technology to open platforms, thus indicating that the benefits of standardizing a preferred technology outweigh the temptation to (p. 44) free-ride in those cases. Where SSOs encourage horizontal openness, that should encourage innovation in complementary markets by expanding the addressable market or installed base. And while standards can reduce the scope for horizontal differentiation in the market for a focal component, increased competition may stimulate the search for extensions and other “vertical” quality enhancements, as emphasized in Bresnahan's (2002) analysis of divided technical leadership in the personal computer industry and the quality-ladder model of Acemoglu et al. (2010). Finally, in horizontally closed platforms, SSOs may encourage innovation by enabling commitments to vertical openness.13 In particular, when a platform leader controls some bottleneck resource, small entrants may fear higher access prices or other policy changes that would capture a share of their innovation rents. Platform leaders might solve this hold-up problem by using SSOs to commit to ex post competition (see generally Farrell and Gallini, 1988). For instance, Xerox researchers used an SSO to give away the Ethernet protocol, and Microsoft took the same strategy with ActiveX (Sirbu and Hughes, 1986; Varian and Shapiro, 1998, 254). One important way SSOs address potential hold-up problems is by requiring firms to disclose essential patents and Page 8 of 20 Four Paths to Compatibility to license them on reasonable and nondiscriminatory (RAND) terms. These policies seek to prevent patent holders from demanding royalties that reflect coordination problems and the sunk costs of implementation, as opposed to the way that well-informed ex ante negotiation would reflect benefits of their technology over the next best solution.14 Although the “reasonable” royalty requirement can be hard to enforce, and SSOs cannot protect implementers from nonparticipating firms, these intellectual property policies are nevertheless an important method for platform leaders to commit to vertical openness. In summary, standard setting organizations are a heterogeneous set of institutions linked by their use of the consensus process. This process emphasizes technical performance, and may select for high-quality standards, but can also produce lengthy delays when participants disagree. SSOs also provide a forum for collaborative innovation, and a way for platform leaders to commit to vertical openness in order to promote market entry and complementary innovation. 3.3. Imposing a Standard A third path to coordination is for someone with sufficient clout to simply impose a standard. This dominant player might be a platform leader, a large customer or complementer, or a government agency. A potential advantage of coordination by fiat is speed. In particular, dictators can avoid or resolve deadlocks that emerge in both standards wars and SSOs. Systemwide architectural transitions may also be easier when a de facto platform leader internalizes the benefits of a “big push” and is therefore willing to bear much of the cost. However, since dictators are not always benevolent or capable of spotting the best technology, ex post competition, innovation incentives, and technical quality will often depend on who is in charge. (p. 45) Platform leaders often dictate standards for vertical interoperability. For example, AT&T historically set the rules for connecting to the US telephone network, and IBM has long defined the interfaces used in the market for plug-compatible mainframes. More recently, Apple has maintained tight control over new applications for its iPhone/iPad platform. In principle, platform leaders should have an incentive to use their control over key interfaces so as to organize the supply of complements efficiently. However, fears that incumbent monopolists will block entry or hold up complementary innovators often lead to calls for policy makers to intervene in support of vertically open interfaces.15 Farrell and Weiser (2003) summarize arguments for and against mandatory vertical openness, and introduce the term internalizing complementary externalities (ICE) to summarize the laissez faire position that a platform leader has socially efficient incentives. When ICE holds, a platform leader chooses vertical openness to maximize surplus; open interfaces promote entry and competition in complementary markets, while closed interfaces encourage coordination and systemic innovation. However, Farrell and Weiser note multiple exceptions to the ICE principle, making it difficult to discern the efficiency of a platform leader's vertical policies in practice. For example, a platform sponsor may inefficiently limit access if it faces regulated prices in its primary market; if control over complements is a key tool for price discrimination; if it has a large installed base; or if a supply of independent complements would strengthen a competing platform.16 In addition to platform leaders, large customers or complementers can act as de facto standard setters. For instance, WalMart played an important role in the standardization for radio frequency identification (RFID) chips by committing to a particular specification. Similarly, movie studios played a significant role in resolving the standards war between Blu-ray and HD-DVD. And in some cases, the pivotal “customer” is actually a user group, as when Cable Labs—a consortium of broadcasters—developed the data over cable service interface specification (DOCSIS) protocol for cable modems. The interests of large complementers and direct customers are often at least loosely aligned with those of endusers, to the extent that their own competitive positions are not threatened. Thus consumers may well benefit from choices made by those powerful players. However, even well-informed powerful players may find it useful to gather information within an SSO before making a decision. Farrell and Simcoe (2012) model this hybrid process, and find that it often outperforms both uninformed immediate random choice and an SSO-based screening process that lacks a dominant third-party. Government is a third potential dictator of standards. In some cases, the government exerts influence as a large Page 9 of 20 Four Paths to Compatibility customer. For example, the US Office of Management and Budget Circular A-119 encourages government agencies to use voluntary consensus standards. And in response to requests from the European Union, Microsoft submitted its Open Office XML file formats to ISO. More controversially, governments may use regulatory authority to promote a standard. For example, the US Federal Communication Commission coordinated a switch from (p. 46) analog (NTSC; named for the National Television System Committee) to digital (ATSC; named for the Advanced Television System Committee) television broadcasting. Sometimes support for standards is even legislated, as in the 2009 stimulus package, which contains incentives for physicians to adopt standardized electronic medical records (but does not take a position on any specific technology). In general, government involvement can be relatively uncontroversial when there are large gains from coordination and little scope for innovation or uncertainty about the relative merits of different solutions. For instance, it is useful to have standards for daylight-saving time and driving on the right side of the road. Government involvement may also be appropriate when private control of an interface would lead to extreme market power primarily because of severe coordination problems as opposed to differences in quality. But government intervention in highly technical standard-setting processes can pose problems, including lack of expertise, regulatory capture, and lock-in on the government-supported standard. 3.4. Converters and Multihoming Converters, adapters, translators, and multihoming are ways to reduce the degree or cost of incompatibility. For example, computers use a wide variety of file formats to store audio and video, but most software can read several types of files. Econometricians use translation programs, such as Stat Transfer, to share data with users of different statistical software. Even the Internet's core networking protocols arguably function as a cross-platform converter: as long as all machines and networks run TCP/IP, it is possible to connect many different platforms and applications over a wide variety of physical network configurations. One benefit of using converters to achieve compatibility is that no single party incurs the full costs of switching. Rather, everyone can choose their preferred system but can also tap into another platform's supply of complements, albeit at a cost and perhaps with some degradation. Since translators need not work in both directions, development costs are typically incurred by the party who benefits, or by a third party who expects to profit by charging those who benefit. And converters can avert the long deadlocks that may occur in a standards war or an open-ended negotiation, since there is no need to agree in advance on a common standard: each platform simply publishes its own interface specifications and lets the other side build a converter (assuming unanimous support for the converter-based solution). Sometimes users or complementers may join several platforms; such multihoming can resemble a converter solution. For example, most retailers accept several payment card systems, so consumers can pick one and for the most part not risk being unable to transact. Corts and Lederman (2009) show that video game developers increasingly multihome, and argue that multiplatform content explains declining concentration in the console market over time. And instead of seeking a common standard for all computer cables, most machines offer a variety of sockets (p. 47) to accommodate different connectors such as USB, SCSI, HDMI, and Ethernet. Multihoming preserves platform variety and may align the costs and benefits of horizontal compatibility. However, dedicated converters or coordination on a single platform become more efficient as the costs of platform adoption increase. Of course, multihoming or converters cannot eliminate conflicting interests, and can open new possibilities for strategic behavior. For example, firms may seek an advantage by providing converters to access a rival's complements while attempting to isolate their own network. Atari tried this strategy by developing a converter to allow its users to play games written for the rival Nintendo platform. However, Nintendo was able to block Atari's efforts by asserting intellectual property based on an encryption chip embedded in each new game (Shapiro and Varian, 1998). Firms may use one-way converters to create market power on either side of a vertical interface. MacKie-Mason and Netz (2007) suggest that Intel pursued this strategy by including a “host controller” in the USB 2.0 specification, and allowing peripheral devices to speak with the host-controller, but delaying the release of information about the link between the host-controller and Intel's chipsets and motherboards. Page 10 of 20 Four Paths to Compatibility Converters can also favor a particular platform by degrading, rather than fully blocking, interoperability. Many computer users will be familiar with the frustrations of document portability, even though most word processors and spreadsheets contain converters that read, and sometimes write, in the file formats used by rival software. Finally, converters may work poorly for technical reasons. This may be particularly salient for vertical interfaces, since allowing designs to proliferate undercuts the benefits associated with modularity and specialization across components. For example, most operating systems do not provide a fully specified interface for third-party hardware (e.g. printers or keyboards), and the “device driver” software that acts as a translator is widely believed to be the most common cause of system failures (Ganapathi et al., 2006). In summary, converters are attractive because they preserve flexibility for implementers. However, in a standards war, firms may work to block converters, as Atari did. Firms may also gain competitive advantage by using converters to manipulate a vertical interface. And even when there is little conflict, dedicated compatibility standards may dominate converters for heavily used interfaces, where performance and scalability are important. 4. Choosing a Path What determines which path to compatibility is followed, or attempted, in a particular case? When will that choice be efficient? While data on the origins of compatibility standards are scant, this section offers some remarks on the selection process.17 (p. 48) Choosing a path to compatibility can itself be a coordination problem, creating an element of circularity in analysis of this choice. We try to sidestep this logical dilemma by grouping platforms into two categories: those with a dominant platform leader, and shared platforms that default to either collective governance (SSOs) or splintering and standards wars. Eisenmann (2008) suggests that this distinction between shared and proprietary platforms emerges early in the technology life cycle, based on firms’ strategic decisions about horizontal openness. In particular, he predicts that platform leaders will predominate in “winner-take-all” markets where network effects are large relative to the demand for variety, multihoming is costly, and the fixed costs of creating a new platform are substantial. This life-cycle perspective of platform governance is consistent with the intriguing (though unsystematic) observation that many technologies settle on a particular path to compatibility, even a specific agency, and adhere to it over time. For example, the ITU has managed international interoperability of telecommunications networks since 1865, and JEDEC has been the dominant SSO for creating open standards for semiconductor interoperability (particularly in memory chips) since 1968. Likewise, for products such as operating systems and video game consoles, proprietary platform leadership has been the dominant mode of coordination across several generations of technology. Nevertheless, there are several well-known cases of dominant firms losing de facto control over a platform. The most famous example is IBM and the personal computer architecture. Other examples include the demise of microcomputing incumbents like Digital Equipment; Google replacing AltaVista as the dominant search engine; Microsoft's well-documented struggles to adapt to the Internet; and the ongoing displacement of the Symbian cellular phone operating system by alternatives from Research in Motion (Blackberry), Apple (iPhone), and Google (Android). Bresnahan (2001) suggests that a key condition for such “epochal” shifts in platform leadership is disruptive technical change at adjacent layers of the larger system, since it is hard to displace a platform leader through direct horizontal competition when network effects are strong. Although this observation certainly accords with the facts of well-known cases, it is not very amenable to formal testing given the infrequent nature of such major shifts. 4.1. Selection and Efficiency When there is a clear platform leader, the ICE principle suggests that leader will have an incentive to choose an efficient coordination process regarding vertical compatibility. For example, the platform leader might delegate standard-setting activities to an SSO when fears of hold-up impede complementary innovation, but impose a standard when SSO negotiations deadlock. Page 11 of 20 Four Paths to Compatibility Unfortunately, the ICE principle is subject to many caveats, bringing back questions about whether a platform leader chooses a particular path for its efficiency or for other reasons such as its impact on ex post competition. For instance, Gawer and (p. 49) Henderson (2007) use Intel's decision to disseminate USB as an example of ICE, while MacKie-Mason and Netz (2007) argue that Intel manipulated USB 2.0 to gain a competitive advantage. Where one study emphasizes the initial decision to give up control over a technology, the other emphasizes the use of one-way converters to exclude competitors and gain lead-time advantages in complementary markets. These competing USB narratives highlight the difficulty of determining a platform leader's motives. Without a platform leader, it is less clear why the private costs and benefits of choosing an efficient path to compatibility would be aligned. If firms are ex ante symmetric and commit to a path before learning the merits of competing solutions, they would have an ex ante incentive to choose the efficient mechanism. But standard setting is typically voluntary, and firms do not commit to abide by consensus decisions. Thus, when asymmetries are present, or emerge over time, firms may deviate to a path that favors their individual interests. While these deviations from collective governance may lead to the “forking” of standards, they do not necessarily block the SSO path, and in some cases the remaining participants can still achieve converter-based compatibility. Some observers suggest that this chaotic situation can deliver the virtues of both decentralized adoption and collective choice. For example, Greenstein (2010) argues that a proliferation of SSOs combined with widespread independent technical experimentation is a sign of “healthy standards competition” on the commercial Internet. This optimistic view emphasizes the virtues of “standards swarms.” When network effects are weak (as the absence of a platform leader might sometimes suggest), and substantial market or technological uncertainty exists, decentralized choice can identify promising standards for a particular niche, with SSOs emerging to facilitate coordination as needed. Unfortunately, there is no guarantee that mixing decentralized adoption with SSOs captures the benefits and avoids the costs of either path in isolation. In particular, either path may lead to a stalemate, and when decentralized adoption is the outside option there is always a danger of stranded investments or selecting the wrong system. A second optimistic argument holds that new ways to govern shared technology platforms will arise in response to market pressures and technological opportunities. For example, Cargill (2001) and Murphy and Yates (2009) claim that accredited SDOs lost market share to small consortia during the 1980s and 1990s because the SDOs’ ponderous decision-making procedures were ill-matched to rapid ICT product lifecycles (see also Besen and Farrell 1991). Smaller and less formal organizations might work faster by relaxing the definition of consensus, taking advantage of new technologies for collaboration, and allowing competing factions to work in isolation from one another. Berners-Lee and Fischetti (1999) cite delays at the IETF as a primary motive for creating the World Wide Web Consortium (W3C), and Figure 2.2 shows that consortia are indeed on the rise.18 Figure 2.2 The Growth of Consortia. Notes:Figure shows the cumulative number of new consortia founded during each five-year period, based on the authors' analysis of the list of ICT consortia maintained by Andrew Updegrove, and published at www.consortiuminfo.org.. While this evolutionary hypothesis is intriguing, it is not obvious that organizational experimentation and competition will evolve an efficient path to compatibility. Simcoe (2012) shows that consortia still experience coordination delays when participants have conflicting interests over commercially significant technology. (p. 50) And the proliferation of SSOs also increases the potential for forum shopping, as emphasized by Lerner and Tirole. We view competition between SSOs as a promising topic for further research.19 At present, it remains unclear Page 12 of 20 Four Paths to Compatibility whether the current proliferation of organizational models for SSOs is the outcome of, or part of, an evolutionary process, or simply confusion regarding how best to organize a complex multilateral negotiation. 4.2. Hybrid Paths Although markets, committees, converters, and dictators offer distinct paths to compatibility, they can sometimes be combined. For example, standards wars may be resolved through negotiations at an SSO, or the intervention of a dominant firm; and slow SSO negotiations may be accelerated by an agreement to use converters, or by evidence that the market is tipping toward a particular solution. Farrell and Saloner (1988) model a hybrid coordination process that combines markets and committees. In their model, the hybrid path combines the virtues of standards wars and SSOs without realizing all of the costs. In particular, the market works faster than an SSO, while the committee reduces the chance of inefficient splintering. The IETF's informal motto of “rough consensus and running code” reflects a similar logic. By emphasizing “running code,” the IETF signals that firms should not wait for every issue to be resolved within a committee, and that some level of experimentation is desirable. However, it remains important to achieve at least “rough consensus” before implementation. Synergies of hybrid style can also occur between SSOs. Independent firms and small consortia often work to preestablish a standard, before submitting it to an accredited SDO for certification. For example, Sun Microsystems used ISO's (p. 51) publicly accessible specification (PAS) process to certify the Java programming language and ODF document format (Cargill 1997). Similarly, Microsoft used ISO's fast-track procedures to standardize its Open Office XML document formats. As described above, platform leaders may value SDO certification if it provides a credible signal of vertical openness that attracts complementary innovators. However, critics claim that fast-track procedures can undermine the “due process” and “balance of interest” requirements that distinguish SDOs from consortia, leading users or complementers to adopt proprietary technology out of a false sense of security. A second hybrid path to compatibility occurs when participants in a standards war use converters to fashion an escape route. For example, the 56K modem standards war was resolved by adopting a formal standard that contained elements of competing systems. Converters can also reduce the scope of conflicting interests within an SSO, especially when participants adopt a “framework” that falls short of full compatibility. An alternative escape route (and third hybrid path) relies on a dictator to break deadlocks within an SSO. Farrell and Simcoe (2012) analyze a model of consensus standard setting as a war of attrition, in which a poorly informed but neutral third party can break deadlocks by imposing a standard. They find that this hybrid process will often (but not always) outperform an uninterrupted screening process, or an immediate uninformed choice. In practice, there are many examples of a dominant player intervening to accelerate a consensus process, such as the case of DOCSIS (cable modems) or electronic health records, both mentioned above. Thus, informally, there are some reasons to hope that a standards system with many paths to compatibility will perform well. Platform leaders often have an incentive to choose the efficient path, and a greater variety of “pure” paths means more options to choose from. SSOs may evolve in response to market pressures and technological opportunities. And both theory and practical observation suggest that many paths to compatibility can be combined in complementary ways. At this point in our understanding, however, any optimism should be very cautious. Various exceptions to the ICE principle show that platform leaders may weigh efficiency against ex post competition when choosing a path to compatibility. It is not clear when competition among SSOs will lead to more efficient institutions, as opposed to increased forum shopping and technology splintering. And while hybrid paths can work well, they highlight the complex welfare trade-offs among the probability of coordination, the costs of negotiation, and the implications for ex post competition and innovation. 5. CONCLUSIONS Compatibility standards can emerge through market competition, negotiated consensus, converters or the actions of a dominant firm. These four paths to compatibility have different costs and benefits, which depend on whether a Page 13 of 20 Four Paths to Compatibility (p. 52) particular standard promotes vertical or horizontal interoperability, the presence of an installed base or proprietary complements, firms’ sunk investments in alternative designs, and the distribution of intellectual property rights. When choosing a path to compatibility, there are trade-offs between the probability of coordination, expected costs in time and resources, and the implications for ex post competition and innovation. There is an argument that a platform leader will internalize these costs and benefits and choose the socially efficient path to compatibility. But that argument has many exceptions. Others argue that decentralized experimentation with different technologies, loosely coordinated by a combination of markets and SSOs, will typically produce good outcomes. However, it is hard to predict how far competition among SSOs leads them toward optimal policies, or how reliably standards wars select the superior platform. Amid these complex questions, there is certainly scope for beneficial government involvement, whether as a large end-user, a regulator, or a third party with technical expertise. But direct government intervention in highly technical standard-setting processes can pose problems, including lack of expertise, regulatory capture, and lockin on government-supported standards. Viewing the economic literature on compatibility standards in terms of our four broad paths also suggests several research opportunities. First, there are very few data on the relative market share of these alternative paths. Thus, it is unclear whether economists have focused on the most important or most common modes of organizing the search for compatibility, or merely the routes they find most interesting. Our impression is that standards wars and platform leaders have received more academic attention than have SSOs and converters. Possibly this is because the former paths are replete with opportunities for interestingly strategic, yet familiarly market-based, competitive strategies, while the latter options lead to less tractable or more foreign questions of social choice and bargaining. A second topic for research is the selection of a path to compatibility, particularly in the early stages of a technology life cycle. Many studies assume either a standards war or a platform leader (who might delegate the choice of standards for vertical compatibility to an SSO). But we know little about how the rules for collective governance of a shared platform emerge or evolve over time. And there is not much research on forum shopping by technology sponsors, or the nature and effects of competition among SSOs. Developing a better understanding of how a particular path is chosen represents a crucial first step toward quantifying the cost-benefit tradeoffs across paths (unless the assignment is random), and adjudicating debates over the efficiency of the selection process. Finally, there is an opportunity to examine interactions among the four paths to compatibility. Despite some first steps toward modeling “hybrid” paths, there is no general theory and very little empirical evidence on who chooses the mechanism(s) and how, or on whether the four paths tend to complement or interfere with one another. References Acemoglu, D., Gancia, G., Zilibotti, F., 2010. Competing Engines: of Growth: Innovation and Standardization. NBER Working Paper 15958. Arthur, W.B., 1989. Competing Technologies, Increasing Returns, and Lock-In by Historical Events. Economic Journal 97, pp. 642–665. Augereau, A., Greenstein, S., Rysman, M., 2006. Coordination versus Differentiation in a Standards War: 56K modems. Rand Journal of Economics 34(4), pp. 889–911. (p. 55) Baldwin, C.Y., Clark, K.B., 2000. Design Rules. Vol. 1: The Power of Modularity. MIT Press. Bekkers, R., Duysters, G., Verspagen, B., 2002. Intellectual Property Rights, Strategic Technology Agreements and Market Structure: The Case of GSM. Research Policy 31, pp. 1141–1161. Berners-Lee, T., Fischetti, M., 1999. Weaving the Web: The Original Design and Ultimate Destiny of the World Wide Web by Its Inventor. San Francisco: HarperSanFrancisco. Page 14 of 20 Four Paths to Compatibility Besen, S. M., Farrell, J., 1991. The Role of the ITU in Standardization: Pre-Eminence, Impotence, or Rubber Stamp? Telecommunications Policy 15(4), pp. 311–321. Besen, S. M., Farrell, J., 1994. Choosing How to Compete—Strategies and Tactics in Standardization. Journal of Economic Perspectives 8(2), pp. 117–131. Besen, S.M., Saloner, G., 1989. The Economics of Telecommunications Standards. In: R. Crandall, K. Flamm (Eds.), Changing the Rules: Technological Change, International Competition, and Regulation in Telecommunications. Washington DC, Brookings, pp. 177–220. Biddle, B., White, A., Woods, S., 2010. How Many Standards in a Laptop? (And Other Empirical Questions). Available at SSRN: http://ssrn.com/abstract=1619440. Boudreau, K., 2010. Open Platform Strategies and Innovation: Granting Access versus Devolving Control. Management Science 56(10), pp. 1849–1872. Bresnahan, T., 2002. The Economics of the Microsoft Case. Stanford Law and Economics Olin Working Paper No. 232. Available at SSRN: http://ssrn.com/abstract=304701. Bresnahan, T., Greenstein, S., 1999. Technological Competition and the Structure of the Computer Industry. Journal of Industrial Economics 47(1), pp. 1–40. Bresnahan, T., Yin, P.-L., 2007. Standard Setting in Markets: The Browser War. In: S. Greenstein, Stango, V. (Eds.), Standards and Public Policy, Cambridge University Press. Cabral, L., Kretschmer, T., 2007. Standards Battles and Public Policy. In: S. Greenstein, Stango, V. (Eds.), Standards and Public Policy, Cambridge University Press. Cabral, L., Salant, D., 2008. Evolving Technologies and Standards Regulation. Working Paper. Available at SSRN: http://ssrn.com/abstract=1120862. Cargill, C., 2001. Evolutionary Pressures in Standardization: Considerations on ANSI's National Standards Strategy. Testimony Before the U.S. House of Representatives Science Committee. (http://www.opengroup.org/press/cargill_13sep00.htm). Cargill, C., 2002. Intellectual Property Rights and Standards Setting Organizations: An Overview of Failed Evolution. Available at: www.ftc.gov/opp/intellect/020418cargill.pdf. Cargill, C. F., 1989. Information Technology Standardization: Theory, Process, and Organizations. Bedford, Mass, Digital Press. Cargill, C. F., 1997. Open Systems Standardization: A Business Approach. Upper Saddle River, N.J., Prentice Hall PTR. Chiao, B., Lerner, J., Tirole, J., 2007. The Rules of Standard Setting Organizations: An Empirical Analysis. RAND Journal of Economics 38(4), pp. 905–930. Corts, K., Lederman, M., 2009. Software Exclusivity and the Scope of Indirect Network Effects in the U.S. Home Video Game Market. International Journal of Industrial Organization 27(2), pp. 121–136. (p. 56) David, P. A., Greenstein, S., 1990. The Economics of Compatibility Standards: An Introduction to Recent Research. Economics of Innovation and New Technology 1(1), pp. 3–42. David, P. 1988. Clio and the Economics of QWERTY. American Economic Review 75, pp. 332–337. DeLacey, B., Herman, K., Kiron, D., Lerner, J., 2006. Strategic Behavior in Standard-Setting Organizations. Harvard NOM Working Paper No. 903214. Dranove, D., Gandal, N., 2003. The DVD vs. DIVX Standard War: Empirical Evidence of Network Effects and Preannouncement Effects. The Journal of Economics and Management Strategy 12(3), pp. 363–386. Page 15 of 20 Four Paths to Compatibility Dranove, D., Jin, G., 2010. Quality Disclosure and Certification: Theory and Practice. Journal of Economic Literature 48(4), pp. 935–963. Eisenmann, T. 2008. Managing Proprietary and Shared Platforms. California Management Review 50(4). Eisenmann, T., Barley, L., 2006. Atheros Communications. Harvard Business School, Case 806–093. Eisenmann, T., Parker, G., Van Alstyne, M., 2009. Opening Platforms: When, How and Why? In: A. Gawer (Ed.), Platforms, Markets and Innovation. Cheltenham, UK and Northampton, Mass.: Edward Elgar Publishing. European Commission 2010. Guidelines on the Applicability of Article 101 of the Treaty on the Functioning of the European Union to Horizontal Co-operation Agreements. Available at: http://ec.europa.eu/competition/consultations/2010_horizontals/guidelines_en.pdf Farrell, J., 2007. Should Competition Policy Favor Compatibility? In: S. Greenstein, Stango, V. (Eds.), Standards and Public Policy, Cambridge: Cambridge University Press, pp. 372–388. Farrell, J., Gallini, N., 1988. Second-Sourcing as a Commitment: Monopoly Incentives to Attract Competition. The Quarterly Journal of Economics 103(4), pp. 673–694. Farrell, J., Hayes, J., Shapiro, C., Sullivan, T., 2007. Standard Setting, Patents and Hold-Up. Antitrust Law Journal 74, pp. 603–670. Farrell, J., Klemperer, P., 2007. Coordination and Lock-in: Competition with Switching Costs and Network Effects. In: M. Armstrong, Porter, R.H. (Eds.), Handbook of Industrial Organization (Volume 3), Elsevier, pp. 1967–2056. Farrell, J., Saloner, G., 1988. Coordination through Committees and Markets. Rand Journal of Economics 19(2), pp. 235–252. Farrell, J., Simcoe, T., 2012. Choosing the Rules for Consensus Standardization. The RAND Journal of Economics, forthcoming. Available at: http://people.bu.edu/tsimcoe/documents/published/ConsensusRules.pdf Farrell, J., Shapiro, C., 1992. Standard Setting in High-Definition Television. Brookings Papers on Economic Activity, pp. 1–93. Farrell, J., Weiser, P., 2003. Modularity, Vertical Integration, and Open Access Policies: Toward a Convergence of Antitrust and Regulation in the Internet Age. Harvard Journal of Law and Technology 17(1), pp. 85–135. Furman, J., Stern, S., 2006. Climbing Atop the Shoulders of Giants: The Impact of Institutions on Cumulative Research. NBER Working Paper No. 12523. Gallagher, S., West, J., 2009. Reconceptualizing and Expanding the Positive Feedback Network Effects Model: A Case Study, Journal of Engineering and Technology Management 26(3), pp. 131–147. (p. 57) Gawer A., Henderson, R., 2007. Platform Owner Entry and Innovation in Complementary Markets: Evidence from Intel, Journal of Economics & Management Strategy 16(1), pp. 1–34. Ganapathi, A., Ganapathi, V., Patterson, D., 2006. Windows XP Kernel Crash Analysis. Proceedings of the 20th Large Installation System Administration Conference, pp. 149–159. Greenstein, S., 2010. Glimmers and Signs of Innovative Health in the Commercial Internet. Journal of Telecommunication and High Technology Law 8(1), pp. 25–78. Henderson, R., 2003. Ember Corp.: Developing the Next Ubiquitous Networking Standard. Harvard Business School, Case 9–703–448. Henderson, R., Clark, K., 1990. Architectural Innovation: The Reconfiguration of Existing Product Technologies and the Failure of Established Firms. Administrative Science Quarterly 35(1), pp. 9–30. Intellectual Property Owners Association, 2009. Standards Primer: An Overview of Standards Setting Bodies and Patent-Related Issues that Arise in the Context of Standards Setting Activities. Section 16. Available at: Page 16 of 20 Four Paths to Compatibility standardslaw.org/seminar/class-2/excerpts-from-ipo-standards-primer/. Kaplan, J., 1986. Startup: A Silicon Valley Adventure. New York: Penguin. Katz, M. L., Shapiro, C., 1985. Network Externalities, Competition and Compatibility. American Economic Review 75, pp. 424–440. Liebowitz, S. J., Margolis, S., 1990. The Fable of the Keys, Journal of Law and Economics 33(1), pp. 1–25. Lemley, M., 2002. Intellectual Property Rights and Standard Setting Organizations. California Law Review 90, pp. 1889–1981. Lerner, J., Tirole, J., 2006. A Model of Forum Shopping. American Economic Review 96(4), pp. 1091–1113. Lerner, J., Tirole, J., Strojwas, M., 2003. Cooperative Marketing Agreements Between Competitors: Evidence from Patent Pools. NBER Working Papers 9680. Majoras, D. 2005. Recognizing the Pro-competitive Potential of Royalty Discussions in Standard-Setting. Stanford University Standardization and the Law Conference. September 23, 2005. Available at: http://www.ftc.gov/speeches/majoras/050923stanford.pdf. Mackie-Mason, J., Netz, J., 2007. Manipulating Interface Standards as an Anticompetitive Strategy. In: S. Greenstein, Stango, V. (Eds.), Standards and Public Policy, Cambridge University Press. Matutes, C., Regibeau, P., 1988. “Mix and Match”: Product Compatibility without Network Externalities. RAND Journal of Economics 19(2), pp. 221–234. Murphy C., Yates, J., 2009. The International Organization for Standardization (ISO). New York: Routledge. Murray, F., Stern, S., 2007. Do Formal Intellectual Property Rights Hinder the Free Flow of Scientific Knowledge? An Empirical Test of the Anti-Commons Hypothesis. Journal of Economic Behavior and Organization 63(4), pp. 648– 687. Parker, G., Van Alstyne, M., 2005. Two-Sided Network Effects: A Theory of Information Product Design. Management Science 51(10), pp. 1494–1504. Rochet, J.-C., Tirole, J., 2003 Platform Competition in Two-sided Markets Journal of the European Economic Association 1(4), pp. 990–1029. Russell, A. L., 2006. Rough Consensus and Running Code and the Internet-OSI Standards War. IEEE Annals of the History of Computing 28(3), pp. 48–61. (p. 58) Rysman, M., 2009. The Economics of Two-Sided Markets. Journal of Economic Perspectives 23(3), pp. 125–143. Rysman, M., Simcoe, T., 2008. Patents and the Performance of Voluntary Standard Setting Organizations, Management Science 54(11), pp. 1920–1934. Shapiro, C., 2001. Navigating the Patent Thicket: Cross Licenses, Patent Pools, and Standard Setting. In: A. Jaffe, J. Lerner, S. Stern (Eds.), Innovation Policy and the Economy (Volume 1), MIT Press, pp. 119–150. Shapiro, C., Varian, H.R., 1998. Information Rules: A Strategic Guide to the Network Economy. Boston, Mass., Harvard Business School Press. Simcoe, T., 2007. Explaining the Increase in Intellectual Property Disclosure. In: Standards Edge: The Golden Mean. Bolin Group. Simcoe, T., 2012. Standard Setting Committees: Consensus Governance for Shared Technology Platforms. American Economic Review 102(1), 305–336. Simcoe, T., Graham, S. J., Feldman, M., 2009. Competing on Standards? Entrepreneurship, Intellectual Property and Platform Technologies. Journal of Economics and Management Strategy 18(3), pp. 775–816. Page 17 of 20 Four Paths to Compatibility Spence, M., 1975. Monopoly, Quality, and Regulation. Bell Journal of Economics 6(2), pp. 417–429. Varney, C. A., 2010. Promoting Innovation Through Patent and Antitrust Law and Policy, Remarks prepared for join USPTO FTC Workshop on the Intersection of Patent Policy and Antitrust Policy. May 26, 2010. Available at: http://www.justice.gov/atr/public/speeches/260101.htm Weiss, M. B., Sirbu, M., 1990. Technological Choice in Voluntary Standards Committees: An Empirical Analysis. Economics of Innovation and New Technology 1(1), pp. 111–134. Weiss, M., Toyofuku, R., 1996. Free-ridership in the Standards-setting Process: The Case of 10BaseT, StandardView 4(4), pp. 205–212. West, J., 2007. The Economic Realities of Open Standards: Black, White and Many Shades of Gray. In: S. Greenstein, Stango, V. (Eds.), Standards and Public Policy, Cambridge University Press, pp. 87–122. Weyl, G., 2010. A Price Theory of Multi-Sided Platforms. American Economic Review 100(4), pp. 1642–1672. Notes: (1.) A user is said to “multihome” when it adopts several incompatible systems and can thus work with others on any of those systems. (2.) See David and Greenstein (1990) or Shapiro and Varian (1998) for a review of the early literature on network effects, and Rysman (2009) for a review of the nascent literature on two-sided markets. (3.) Distinguishing between horizontal and vertical compatibility may help illuminate the often murky concept of open standards. Cargill (1997) suggests that the term “open” has become “an icon to conveniently represent all that is good about computing,” so when conflicts emerge, all sides claim support of open standards. End-users typically define “open” in horizontal terms, since they seek a commitment to future competition at the platform level. Platform leaders typically emphasize vertical compatibility, which grants access to (but not control over) proprietary technology. Meanwhile, standards mavens call a technology open if it has been endorsed by an accredited SSO, and open-source advocates focus on free access to the underlying code. (4.) Shapiro and Varian (1998) provide many other examples, and West (2007) contains a lengthy list of standards wars. Farrell and Klemperer (2007) review the economic theory. (5.) Farrell and Saloner (1988) model sequential technology adoption with network effects and show how outcomes may depend on users’ initial beliefs. The book StartUp (Kaplan, 1986, Ch. 9) provides an entertaining account of the battle for expectations, and the strategic use of backward compatibility, in pen-based computer operating systems. (6.) Cabral and Kretschmer (2007) even suggest that when tipping toward an inferior technology would be very costly, optimal government policy may be to prolong a standards war so participants can gather more information. (7.) Interestingly, many of the well-known standards wars that do result in a “fight to the death” involve media formats. (8.) A list of roughly 550 SSOs is available at www.consortiuminfo.org. Cargill (2002) and the Intellectual Property Owners Association (2009) suggest classification schemes. (9.) These acronyms stand for International Telecommunications Union (ITU), International Organization for Standards (ISO), Internet Engineering Task Force (IETF), and World Wide Web Consortium (W3C). For ICT standards, ISO and the International Electrotechnical Commission (IEC) collaborate through a group called the Joint Technical Committee (JTC 1). Murphy and Yates (2009) describe the relationship between these national standards bodies and the global standards system administered by ISO. (10.) Annex 1 of the World Trade Organization's Technical Barriers to Trade Agreement makes explicit reference to the ISO/IEC Guidelines for SDO procedure. In the United States, OMB Circular A-119 gives preference to SDOs in Page 18 of 20 Four Paths to Compatibility federal government purchasing, and the Standards Development Act of 2004 (HR 1086) grants certain antitrust exemptions. Speeches by Majoras (2005) and Varney (2010), and also the European Commission (2010) antitrust guidelines on horizontal co-operation provide some assurance to SSOs regarding open discussion of royalty rates. (11.) Compatibility standards make up roughly 43 percent of the total stock of American National Standards, with much of the other half related to performance measurement and safety. Thus the ICT sector's share of standards production exceeds its share of GDP, and even patenting. One explanation is that information technology is, by design, uniquely modular, so the ratio of standards to products is high. For example, Biddle et al (2010) estimated that a typical laptop implements between 250 and 500 different compatibility standards. (12.) The decision to adopt a framework or incorporate vendor-specific options into a standard may be observable, and hence amenable to empirical research, since the work-process and published output of many important SSOs (e.g. 3GPP and the IETF) are publicly accessible. (13.) Furman and Stern (2006) and Murray and Stern (2007) study the impact of “vertical” open access policies on innovation outside of network industries. (14.) Farrell et al. (2007) describe an extensive legal and economic literature on SSO IPR policies. (15.) Calls for mandatory vertical openness often produce fierce debates. For example, prior to the 1956 consent decree, IBM resisted publishing the technical specifications that would allow vendors to offer “plug compatible” mainframes and peripherals. More recent are the debates over “net neutrality” and ISPs’ freedom to charge prices based on the quality-of-service provided to different web sites or Internet applications. (16.) Weyl (2010) offers an alternative price-theoretic analysis of a monopoly that controls access to both sides of a two-sided platform. In his model, the monopoly tariffs exhibit two types of deviation from perfectly competitive pricing: a “classical market power distortion” (which resembles a Lerner markup rule) and a “Spence (1975) distortion” whereby the monopolist internalizes the network benefits to the marginal, as opposed to the average, platform adopter. (17.) One exception to the paucity of data is a paper by Biddle et al. (2010) that identifies 250 compatibility standards used in a typical laptop. The authors worked with Intel to estimate that 20 percent of these standards come from individual companies, 44 percent from consortia and 36 percent from accredited SDOs. (18.) Observing the slowdown at many consortia, Cargill suggests that they too will be supplanted, perhaps by the open-source software development model, or other bottom-up efforts to establish de facto standards. (19.) A parallel literature on voluntary certification programs, reviewed in Dranove and Jin (2010), may offer insights on competition between SSOs that can also be applied to compatibility standards. For instance, they cite several recent studies that examine the proliferation of competing “eco-labeling” initiatives (e.g. Energy Star versus LEED for construction or Sustainable Forest Initiative versus Forest Stewardship Council for lumber). Joseph Farrell Joseph Farrell is Professor of Economics at the University of California, Berkeley. Timothy Simcoe Timothy Simcoe is Assistant Professor of Strategy and Innovation at the Boston University School of Management. Page 19 of 20 Software Platforms Oxford Handbooks Online Software Platforms Andrei Hagiu The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0003 Abstract and Keywords This article presents a detailed view on developments in business strategies in the market of software platforms. The fundamental characteristic of multisided platforms is the presence of indirect network effects among the multiple “sides.” The broader implication of the Brightcove and Salesforce examples is that web-based software platforms (SPs) challenge the notion that software platforms are inherently multisided. The article then describes the platform governance. The emergence of cloud-based platforms and virtualization are increasingly challenging its traditional status as the key multisided software platform (MSSP) in computer-based industries. Multi-sidedness essentially depends on whether the provider of the SP owns a related product or service, whose value is increased by the applications built upon the SP. Governance rules appear to have emerged as an important strategic instrument for SPs. Keywords: software platforms, multisided platforms, Brightcove, Salesforce, platform governance, cloud-based platforms, virtualization 1. Introduction Since the 2006 publication of Evans, Hagiu, and Schmalensee (2006) (henceforth IE for Invisible Engines), software platforms have continued to increase their reach and importance in the information technology sector. Numerous companies, large and small, have developed software platforms in an attempt to stake out preeminent market positions at the center of large ecosystems of applications and users. Apple has expanded from computers and digital music players into smart phones, and its iPhone has become the leading hardware-software platform in the mobile market. Facebook and LinkedIn have turned their respective social networks into software platforms for third-party application developers.1 Google has launched several new software platforms: the Android operating system for smartphones; App Engine and Chrome operating system (based on the Chrome browser) for web-based applications; OpenSocial application programming interfaces (APIs) for online social networks. Salesforce, the leading provider of cloud-based customer relationship management (CRM) software, has turned its product into a software platform for external developers of business applications (not necessarily related to CRM). Even Lexmark, a printer manufacturer, has caught software platform fever: it recently opened up APIs and launched an app store for third-party developers to build applications running on top of its printers.2 This chapter provides a brief overview of what I believe to be some of the key recent developments in the world of software platforms. The technological and economic drivers of value creation by software platforms are unchanged since IE (p. 60) (Moore's Law, the ever-expanding variety of computer-based devices, economies of scale associated with writing reusable blocks of software code, etc.), but some of their manifestations are novel, which should supply interesting topics for economics and management research. There is now a significantly wider variety of software platform types, mostly due to the rise of web (or cloud)-based computing and of a technology called virtualization. As a result, there are intriguing competitive dynamics between these new software platform species and the incumbents (mostly operating systems). Software platform business models have also evolved in interesting ways along several dimensions: the extent of vertical integration, pricing structures, and the strictness of the governance rules placed on their respective ecosystems. The chapter is organized as follows. Section 2 discusses the distinction between one-sided and multisided software platforms: while the most powerful software platforms are indeed multisided, it is quite possible for many software platforms to thrive even in a one-sided state. Section 3 analyzes some of the key changes in software business models that have emerged since the publication of IE, focusing on integration scope, governance, and pricing structures. Section 4 discusses webbased software platforms and virtualization, which are challenging the preeminence of the operating system in the relevant computing stacks from above and from below, respectively. The conclusion summarizes the key points and draws a few implications for further research.3 2. Multisided Versus One-Sided Software Platforms The notion of software platforms that will be used in this chapter is the same as in IE: a software platform is any piece of software that exposes APIs, allowing other applications to be built on top.4 As emphasized in IE, one of the key economic features of software platforms is their potential to become the foundation for multisided business models. In other words, they can help bring together multiple groups of customers, which generally include at least third-party application or content developers and end-users. Sometimes they also include third-party hardware manufacturers (e.g., Android, Windows) or advertisers (e.g., Google's search engine and related APIs). The economic value created by software platforms comes fundamentally from reducing the application developers’ fixed costs by providing basic functionalities that most applications need. In other words, they eliminate the need for each developer to individually “reinvent the wheel.” But these economies of scale are not sufficient to turn software platforms into multisided platforms. Indeed, the fundamental characteristic of multisided platforms is the presence of indirect network effects among the multiple “sides”—or (p. 61) customer groups—served. This requires that Page 1 of 11 Software Platforms the platform is mediating some sort of direct interaction between agents on different sides and that each side benefits when more members on the other side(s) join the same platform. Indirect network effects are thus distinct from direct network effects, which occur within the same customer group: participation by more members on one side creates more value for each individual member on the same side. A classic example of direct network effects at work is the fax machine. More recently, social networks such as Facebook and LinkedIn were primarily built around direct network effects among their members. As mentioned in the introduction, however, they have morphed into two-sided platforms by also attracting third-party application developers.5 Some software platforms are indeed two-sided or multisided. This is the case, for example, with Apple's iPhone, Google's Android, and Microsoft's Windows Mobile: they allow users to access and use thousands of applications created by third-party developers and, vice versa, they enable thirdparty application developers to reach the millions of users who own phones running the corresponding operating system. It is important to note, however, that not all software platforms (henceforth SPs) manage to fulfill their multisided potential and become multisided software platforms (henceforth MSSPs). In fact, the last four years have witnessed a proliferation of one-sided software platforms at various layers of computer-based industries: these SPs create value through economies of scale and specialization but not through indirect network effects. One-sidedness may be the result of conscious choice or failure to attract multiple sides on board. The experience of Brightcove, the creator of a SP for online video, provides a good illustration of the challenges associated with the transition from one-sided SPs to MSSPs. The following subsection is based on Hagiu, Yoffie, and Slind (2007) and the authors’ subsequent interviews with the company. 2.1. Brightcove: Four-Sided Vision, One-Sided Reality (Note: This subsection is based on Hagiu, Yoffie, and Slind (2007) and the authors’ subsequent interviews with the company.) Brightcove was founded in 2005 with the ambition to become a four-sided platform in the rapidly growing market for Internet video. Its vision was to connect content providers (e.g., MTV, Sony Music, Discovery Channel, The New York Times), end-users, advertisers, and web affiliates (i.e., web properties who would want to license video streams and integrate them in their sites). Brightcove would achieve this vision by providing several key pieces: a software platform that would enable content providers to build and publish highquality video streams online, an online site that would aggregate videos and allow users to view them, an advertising network enabling content providers to sell 15 or 30-second spaces in their videos to advertisers and a syndication marketplace enabling third-party affiliated websites to license videos from content providers for publication on the affiliates’ sites. (p. 62) The Brightcove team recognized the complexity involved in building this four-sided platform all at once and reasoned that the first two sides they needed to get on board were content providers and users. After some internal debate, they decided to start courting the content provider side first, in particular premium content providers (e.g., Wall Street Journal, Disney, Discovery, MTV, etc.), through the provision of the Brightcove software platform. Content providers relied on it to create and integrate video in their own websites. Brightcove viewed this merely as the first of two steps needed to establish itself as a two-sided software platform: once it had attracted a critical mass of content providers as customers, it would attract the user side by launching its Brightcove.com site and populating it with Brightcove-powered videos from its content providers. In particular, the management team was explicit that the company's goal was not limited to being “just an enterprise software provider,” that is, selling software tools to premium content providers. Brightcove was very successful in attracting premium content publishers as customers for its software platform: many large companies (media, manufacturing, retail) and organizations (government, not-for-profit) were quick to sign up. And in October 2006, Brightcove did indeed launch its own user destination site. Unfortunately, however, the site was never able to generate significant traction. This was for two main, related reasons. First, Brightcove's content provider-customers were reluctant to supply content for Brightcove to build its own destination site, which they viewed as competing with their own sites for user attention and advertising revenues. Second, because it had to focus most of its limited resources on serving the large, premium content providers, Brightcove had been unable to dedicate any resources to building its brand for end-users (even if it had had the financial resources and management bandwidth to do so, this would have conflicted with its focus on premium content). In particular, it entirely neglected the usergenerated video content functionalities offered at the time by YouTube and other sites such as Revver and MetaCafe. As a result, while Brightcove was deepening its relationship with premium content providers and building a solid stream of revenues, YouTube was acquiring tens of millions of users (without any revenues). By the end of 2006, Brightcove.com was hopelessly behind (in terms of user page views) not just YouTube, but also at least four other video-sharing sites. Consequently, in April 2008 Brightcove decided to shut down its user destination site, as well as its advertising network, and effectively settled on being a one-sided software platform for premium content providers. Currently, the company is quite successful in this business and has been profitable since 2009, but it is hard to imagine it would ever be able to achieve the $1.7 billion valuation achieved by YouTube with its 2006 sale to Google. In an irony readily acknowledged by the company's leaders, Brightcove might have been able to build YouTube if it had decided to focus on the end-user side first. Instead, Brightcove's choice to be primarily a software platform (an attractive option by many measures, particularly given the clear and significant revenue prospects) constrained its ability to turn itself into a multisided business. Media companies were very cautious not to repeat the mistake made by music studios in the early 2000s, when they had empowered and (p. 63) subsequently become overly dependent on Apple's iTunes platform. Meanwhile, YouTube's riskier approach of focusing first on users, with no clear monetization prospects initially, allowed it to make an end run around the large media companies. Whether this latter approach will eventually succeed (by compelling media companies to join, given YouTube's tremendous user base) remains to be seen. 2.2. Salesforce: Two-Sided or One-Sided Software Platform? The distinction between one-sided and multisided SPs can be quite subtle, particularly in the case of web-based SPs. Consider the example of Salesforce. Founded in 1999 by a former Oracle executive, the company established itself as the leading provider of on demand (i.e., web-based) customer relationship management (CRM) software. Starting in 2005, Salesforce turned itself into an SP by exposing APIs enabling external developers to build on-demand business applications. The corresponding APIs and other developer services were named the Force.com platform. Some of the third-party applications built on top of Force.com are complementary to Salesforce's flagship CRM product. They enhance the value of Salesforce's CRM application, therefore creating indirect network effects. Other applications however are entirely independent of CRM and target completely different customers. They do not generate indirect network effects around Force.com or Salesforce CRM because their customers can get access to and use them directly from the developers’ sites through a web browser. So is Force.com a two-sided or a one-sided SP? If the share of CRM-related applications were very small, then it would arguably be one-sided. This suggests that attempting to draw a clear line between one-sided and two-sided SPs might be difficult—and ultimately not very insightful. It is however interesting to consider the distinction from Salesforce's perspective. While building a two-sided SP around its CRM product might sound appealing (it Page 2 of 11 Software Platforms would increase switching costs for its customers and create network effects), the scope of applications that can be built on top of CRM is quite limited. This is probably why the company has chosen to encourage and support the development of on-demand applications unrelated to CRM: while they do not create indirect network effects, the advantage is to offer larger growth prospects for the Force.com SP. Thus, in theory at least, Salesforce could conceivably abandon the CRM application and focus exclusively on being a one-sided SP provider, with no direct relationships to enterprise software customers. The broader implication of the Brightcove and Salesforce examples is that web-based SPs challenge the notion that software platforms are inherently multisided (as we had claimed in chapter 3 in IE). The fundamental reason is the following. In a 100 percent offline world, in which software platforms were tied to specific devices, the SP providers had to ensure both that developers used their APIs and that end-users had those APIs installed on their devices. For instance, Microsoft had to simultaneously convince PC application developers to rely on Windows (p. 64) APIs and users to buy PCs running the Windows operating system (OS). Every application built on top of Windows makes users more likely to adopt Windows and vice versa. For web-based SPs, the user side of the equation is solved by definition: any user with any device running a web browser is in principle able to access the applications built on top of the SP. In this context, whether a web-based SP is one-sided or two-sided depends essentially on whether its provider owns another asset or online property whose value to its users increases as more applications are built on the web-based SP. And there is a priori no reason for which web-based SP providers should possess complementary assets. 3. Multisided Software Platforms Business Models IE analyzed three key strategic decisions faced by MSSPs: vertical scope (i.e., the extent of vertical integration into one or several of the multiple sides); pricing structures (i.e., which pricing instruments to use and on which side or sides to seek profits); and bundling (i.e., which features and functionalities to include in the platform itself). While little has changed in the drivers and manifestations of bundling by software platforms, there have been some interesting recent developments regarding vertical scope and pricing structure decisions by MSSPs. I shall also discuss an additional strategic decision: platform governance, that is, the controls and rules put into place by some software platforms to “regulate” the access and interactions of the multiple sides they serve.6 While this decision was touched upon briefly in IE (chapter 10), it has become much more prominent in recent years, particularly in the context of new, web-based, and mobile-based software platforms such as Facebook, iPhone, and Android. 3.1. Integration and Scope The three “core” potential sides for any SP are unchanged from the ones described in IE (chapter 9): end-users, third-party application or content developers, and third-party hardware and peripheral equipment manufacturers.7 Some Web-based SPs have an additional side—advertisers. 3.1.1. Two-Sided versus Three-Sided Software Platforms One of the most fascinating forms of platform competition is that between software platforms with different degrees of vertical integration into hardware. The best known (described in IE chapter 4) is Microsoft versus Apple in PCs. Its most interesting correspondent today is Apple (iPod, iPad, iPhone) versus Google (Android) (p. 65) in smartphones and tablets. Just like in PCs, Apple runs a two-sided SP with the iPhone (users and third-party app developers) and makes the bulk of its profits on the user side through its high margins on iPhone sales. Google runs Android as a three-sided SP (users, handset makers, and third-party app developers), very similar to Windows in PCs, except that Android is open and free for handset makers to use. Whereas the two SP battles look quite similar (and share one protagonist, Apple), the outcomes will likely be quite different. Windows won the PC battle thanks to the superior strength of its indirect network effects: having original equipment manufacturers (OEMs) as a third side to its platform made a big difference and helped tip the PC market to Windows. Today, the market for smartphone operating systems is quite fragmented: out of 80.5 million smartphones sold in the third quarter of 2010, 37 percent were running on Symbian OS, 26 percent on Android, 17 percent were iPhones, 15 percent were RIM's Blackberry devices, and 3 percent were based on Windows Mobile, the balance coming from Linux and other OSs.8 The focus on Android and the iPhone is deliberate: Symbian, RIM, and Microsoft have been losing market share steadily (they stood at 45 percent, 21 percent, and 8 percent, respectively, in the third quarter of 2009), to the benefit of Android and the iPhone (Android only had a 3.5 percent share in the third quarter of 2009, while the iPhone was already at 17 percent). It is unlikely that the current state of fragmentation in the smartphone OS market will continue, since it creates significant inefficiencies for third-party application developers. But this market is also unlikely to tip to one dominant software platform in the same way the PC market did, for three main reasons. First, third-party software applications seem to add lower value to smartphones relative to PCs. Second and related, there is significantly more scope for differentiation among smartphones than among PCs based on functionality, hardware aesthetics, and the overall quality of the user experience. Third, the other key players in the smartphone ecosystem (mobile operators and handset manufacturers) are unlikely to allow one SP to dominate the smartphone market in the same way that Windows dominates the PC industry. Both Apple's two-sided model and Google's completely horizontal, three-sided model have strengths and weaknesses. The Apple model with its vertical integration into hardware tends to produce better devices overall given that one company can design and coordinate the interdependencies of all necessary components. Google's horizontal model, on the other hand, allows for much faster growth and proliferation of the software platform by leveraging the efforts of multiple hardware makers. Some of the respective weaknesses are illustrated by the experience of the other major smartphone SPs. RIM has largely used the Apple model, having kept complete control over the hardware. It was a pioneer of the smartphone market with its Blackberry line of smartphones, which became a hit with business users. But unlike Apple, RIM was slow in realizing the value of turning its devices into an attractive platform for third-party application developers. The various Blackberry devices were well-designed as standalone products, but they did not share a common software platform, which made them collectively less (p. 66) appealing for developers. There are currently only about 10,000 apps available for Blackberries, compared to over 250,000 for the iPhone and 75,000 for Android. And as a result, during the past year, RIM has been losing both market share and stock market valuation, mostly as a result of heightened competitive pressure from Apple and Google devices.9 At the other end of the spectrum, Microsoft has struggled mightily with its Windows Mobile smartphone OS, despite using essentially the same three-sided strategy that made the success of Windows in PCs. Part of the problem seems to have been Microsoft's failure to either understand or acknowledge the key differences between PCs and smartphones: its Windows Mobile iterations were constantly criticized for being too bulky and slow for mobile phones. Finally, although Symbian was the world's first smartphone OS and is still the leader in market share, it is rapidly losing ground, as mentioned earlier. After more than seven years on the market, Symbian only managed to attract 5,000 applications: by contrast, it took Apple one year to reach 10,000 applications for its iPhone and 3 years to reach 200,000 (cf. Hagiu and Yoffie, 2009). Symbian used a model that can be best described as “semi-vertical integration,” which turned out to be a major handicap. It was majority-owned by Nokia but attempted to become a threesided SP: this did not succeed because other handset makers were reluctant to fully commit to a SP controlled by their largest competitor. In turn, Nokia wasted time and resources trying to persuade potential licensees that it did not fully control Symbian, instead of fully embracing the Apple vertical integration model and optimizing its phones for Symbian (cf. Hagiu and Yoffie, 2009). Page 3 of 11 Software Platforms Consequently, it is impossible to predict, solely based on the choice of integration scope, whether Apple's or Google's model will be more successful in the smartphone OS market. Unlike PCs, however, it is quite possible that both might thrive and coexist. 3.1.2. App Stores Have and Have Nots A second prominent element of SP vertical scope nowadays is the provision of a marketplace, or “app store.” App stores such as Android Marketplace, Apple's iTunes, Sony's PlayStation Store, and others enable end-users of a given SP to find, purchase, and download content and applications for that SP. At the time of IE's publication, there were few SPs that provided their own app stores: some mobile carriers did so through their own portals, the most successful of which was NTT DoCoMo's i-mode in Japan; some digital media SPs, led by Apple's iTunes and RealNetworks’ Real Music Store; and, to some extent, Palm in the PDA market (although Palm had an app store, most applications for Palm OS-powered devices were purchased and downloaded through a third-party intermediary called Handango—cf. Boudreau, 2008). Today, app stores have become widespread. By far the largest number is found in the mobile sector, and they are no longer the exclusive preserve of mobile operators. Instead, the largest and most prominent ones are provided by SP vendors and handset makers: Apple's iPhone App Store, Google's Android Marketplace, (p. 67) RIM's BlackBerry App World, Nokia's Ovi Store (associated with its Symbian S60 software platform), and Microsoft's Windows Mobile Marketplace. And app stores are also proliferating in other computer-based industries, in which they were absent only five years ago. For example, the top three videogame console makers all have launched app stores for the first time during the latest console generation (which started in late 2005): Nintendo's Wii Shop Channel, Microsoft's Xbox Live Marketplace, and Sony's PlayStation Store. Users access these stores directly from their consoles in order to purchase and download games, as well as other digital content, most notably movies. Another example is Samsung recently launching an app store for its Internet-connected TV sets, which runs its Samsung Apps software platform.10 Finally, it is noteworthy that some prominent SPs (e.g., Facebook and LinkedIn) have chosen to dispense with app stores, at least for now. What drives SPs’ decisions whether to vertically integrate into the provision of app stores or not? One would expect SPs whose business models depend on tight controls over the format and nature of third-party applications to provide their own app stores. Clearly, that is the case of Apple with its iTunes and iPhone App Store. It is indeed important to recall that Apple's profits come disproportionately from its hardware sales (iPods and iPhones). By exerting tight and exclusive control over content distribution for those devices, Apple is able to further drive up their value relative to competing devices and increase users’ switching costs—for example, by insisting on very low and uniform pricing for music, by imposing proprietary formats for digital content (music and applications) that make it work only on Apple devices.11 The same applies to videogame consoles.12 Game console-specific app stores have only appeared during the latest generation for two simple reasons. First, in previous generations, videogame console content was almost entirely limited to…videogames. By contrast, all major consoles today act as home computers capable of playing movies, surfing the web, and organizing all kinds of digital content. And second, in previous console generations the vast majority of games were sold as shrink-wrapped software on CDs or cartridges, distributed through brick-and-mortar channels. Today, a rapidly increasing share of console content is distributed through the web, which required the provision of online distribution channels by the console manufacturers themselves. In contrast, SPs whose business models rely on broad ecosystems of partners generally leave the provision of app stores to others. Thus, although Google provides its own Android Marketplace (mostly because no one else provided an Android app store when Android was first launched), it also allows its partner handset makers (e.g., HTC, Motorola) and mobile operators (e.g., Verizon) to build their own Android app stores. As this chapter is being written, there are reports emerging that Amazon and BestBuy are also planning to launch Android app stores.13 Although for Apple control over content distribution (app stores) is important in order to differentiate its devices, for Google, “outsourcing” content distribution is essential in order to allow its partners to differentiate their competing offerings from one another. If Google were to insist on being the sole provider of an Android app store, that would make it significantly harder for two competing (p. 68) handset makers or two competing mobile operators to differentiate their respective Android-based devices or services from one another and, in turn, would make them less likely to adopt the Android SP. 3.2. Governance As defined in Boudreau and Hagiu (2009), managed services provider (MSP) “governance rules” refer to the non-price rules and restrictions that MSP providers put in place in order to “regulate” access to and interactions on their platforms. The only example of tight governance rules discussed in IE was that of videogames consoles, which continue to use security chips in order to restrict the entry of game developers. IE also noted that NTT DoCoMo had imposed some mild governance rules around its i-mode MSSP: it does not exclude anyone, but distinguishes between “official” and “nonofficial” content providers. During the past four years, the choice between tight versus loose governance rules has emerged as a key strategic decision for MSPs in general and MSSPs in particular. And there is significant variation in the governance rules chosen by different MSSPs (cf. Boudreau and Hagiu, 2009). At one end of the spectrum, Apple places tight controls and restrictions over its iPhone app store: every single application must be submitted to Apple for screening, and developers can only use Apple-approved technical formats.14 Similarly, LinkedIn is very selective in its approval process for third-party applications built for its social network. At the other extreme, Google approves almost any application written for Android, while Facebook places no restrictions whatsoever on the apps built on its Platform. At the most basic level, one can think of the choice between tight versus loose governance rules as reflecting a choice between quality versus quantity. More precisely, there are three fundamental ways in which tight governance rules can create value for MSPs (cf. Hagiu, 2009). First, they may help avert “lemons market failures” when there is imperfect information about the quality of at least one side. Second, limiting the entry of and therefore competition among the members of one side of the market can help enhance their innovation incentives by assuring them of the ability to extract higher rents. And third, tight governance rules can be used as a way to incentivize the various sides to take actions or make investments which have positive spillovers on other platform constituents. When it comes to MSSPs, these three considerations may be compounded by strategic ones. For instance, Apple claims that its restrictions on iPhone applications are designed to ensure high-quality applications and weed out the low-quality ones, which is all the more important given the large numbers of applications available and the difficulty consumers might have to tell which ones are good before using them (this story is consistent with the first fundamental driver mentioned above). But industry observers and Apple critics point out that an equally important driver of the restrictions might be Apple's desire to favor Apple's own (p. 69) technologies over competing ones (e.g., Apple's iAdsmobile advertising service over Google's AdMob15). Furthermore, MSSPs (and MSPs generally) do not choose governance rules independently from their other strategic decisions (e.g., pricing structures). There are often significant interdependencies that need to be taken into account. For example, one cannot explain the entry restrictions that Page 4 of 11 Software Platforms videogame consoles continue to place on game developers solely based on the “lemons market failure” argument. Indeed, while weeding out poor quality games may have been the primary reason for which Nintendo created these restrictions in the late 1980s, that can hardly be a concern for console manufacturers today. There are hundreds of specialized magazines and websites that review games, making it highly unlikely that poor quality games would crowd out the good quality ones. Instead, the decision to keep limiting the number of game developers by console manufacturers is closely linked to the way they make money. As extensively documented in IE, profits in the console video-gaming business come from games (sales of first-party games and royalties charged to third-party developers), while consoles are sold at a loss. Free entry by developers for a given console would drive game prices and therefore console profits down (cf. Hagiu and Halaburda, 2009). 3.3. Pricing Structures The fundamental drivers of pricing structure choices by MSPs in general and MSSPs in particular are analyzed in depth in IE (chapter 10) and remain valid. In particular, IE documented that all software platforms with the exception of video-game consoles made the bulk of their profits on the user side and charged nothing to third-party application or content developers.16 In contrast, videogame consoles are sold below cost to users and their manufacturers make their profits from royalties charged to third-party game developers. The explanation of these two contrasting pricing structures relies on three key elements (cf. Hagiu and Halaburda, 2009). First, the “razor-blades” structure of consumer demand for videogames does not hold for other software platforms, such as Windows: the number of PC applications a consumer buys is poorly correlated with his or her value for the overall system. Second, as mentioned above, restricting the entry of game developers results in higher quality games, as well as less intense competition between games on the same console. In turn, this allows console makers to extract more value from game developers. Third, the prevailing pricing structure may to some extent reflect path-dependence, that is, the difficulty for later entrants to reverse the pricing structure that was first established by Nintendo in the late 1980s. There are two notable ways in which MSSP pricing structures have evolved since IE. First, more and more MSSPs charge some form of royalty to the third-party application or content developer side. All providers of SPs for smartphones (Apple, Google, Microsoft, RIM, Symbian) charge their respective app developers a 30 percent cut of their revenues (of course, that amounts to 0 for free apps). (p. 70) Facebook does not charge its third-party app developers for making their products available on its Platform, but in the summer of 2010 it started to levy a 30 percent charge on application developers who decide to accept user payments in Credits, a virtual currency introduced by Facebook for its users. Whether the share of profits derived by these SPs from the developer side relative to the user side becomes substantial remains to be seen. For instance, analysts estimated Apple's 2009 profits coming from iPhone handset sales at roughly $6 billion,17 while a January 2010 estimation of Apple's monthly revenues from the iPhone app store stood at $75 million,18 that is, $900 million on a yearly basis. Even without considering the costs of running the app store, that is a ratio of less than 1 to 6 between profits coming from the developer side (through the app store) and profits coming from the user side (through iPhone sales). Second, many “modern” MSSPs have advertisers as a third or fourth side to their platform and derive substantial revenues from this side. This is to a large extent related to the fact that the majority of the new MSSPs are web-based software platforms. By contrast, advertisers were rarely present on the “classic” MSSPs studied in IE (e.g., PC and PDA operating systems, mobile operator platforms such as i-mode, videogame consoles). The value created for advertisers is quite clear in the case of Facebook: access to the eyeballs of 500 million users, who spend an average of 40 minutes per day on the site.19 It is less clear in the case of a SP like Android, which does not have advertisers as one of its sides—at least not directly. But while advertisers are not directly tied to Android, they are a major customer on Google's other MSP—the search engine—and its major source of revenues. And there are powerful complementarities between Android and Google's search engine: Android-based phones make Google's search engine the default and therefore help drive more user traffic to it, which in turn increases the advertising revenues for the search platform. Furthermore, Google views Android as a strategic weapon designed to preempt the dominance of proprietary software platforms such as Apple's iPhone, RIM's Blackberry, and Microsoft's Windows Mobile—these alternative software platforms may indeed choose to divert mobile search revenues to their respective online properties and away from Google. 4. Multi-Layered Platform Competition One of the novel aspects of competitive dynamics between software platforms today is that they create more and more competition between firms across different layers of the relevant “technology stacks”—as opposed to competition within the same layer. This is because most technology companies recognize that controlling a (multisided) software platform can enable them to extract larger rents from the respective ecosystems in which they play. Indeed, the ecosystems of firms producing computer-based devices and corresponding content and applications create (p. 71) significant economic value for end-users as a whole. At the same time, however, there is intense competition among firms producing the various layers of those systems (chips, hardware, operating system, connectivity service, content, and applications) to extract a larger fraction of that value as profits for themselves.20 In this context, SPs create economic value for all ecosystem participants by reducing the costs of building applications and content for the relevant devices. But the very features that make SPs valuable to end-users and other ecosystem players—namely economies of scale or network effects or both—also endow SPs with significant market power. Over time, successful SPs—particularly MSSPs—create both technological and economic lock-in and end up obtaining tremendous control and influence over all the other layers in the ecosystem. Of course, this type of competitive dynamic is present to some extent in any industry whose products are made of complementary components supplied by different players. But it has become particularly salient in computer-based industries, as the relevant “stacks” have grown increasingly complex and multilayered. The stakes are also higher for software platforms given the tremendous potential value of the indirect network effects created between end-users, application developers, and, where relevant, advertisers. There are essentially two mechanisms through which SP competition across layers arises: (1) companies operating in layers other than the operating system (e.g., hardware, chips, service, content) attempting to provide their own version of the operating system layer; (2) companies attempting to build SPs at layers above or below the operating system. The most common manifestation of the first mechanism is hardware or microprocessor manufacturers deciding to build their own operating systems in order to avoid commoditization and hold-up at the hands of a third-party OS provider. Indeed, the latter scenario prevailed in the PC industry, where Microsoft's Windows MSSP extracted most of the value, leaving very little to OEMs (cf. Yoffie et al., 2004). The desire to prevent a repeat of this scenario in the mobile phone market was the primary driver behind the 1998 founding of Symbian by Nokia and several other prominent handset makers.21 In a similar and more recent move, Samsung decided to create its own Bada operating system for use in its smartphones in 200922 and then, in 2010, extended it into an operating system and corresponding app store for its phones and high-definition television sets, called Samsung Apps.23 Intel itself, Microsoft's closest partner in the PC industry, has recently entered the operating system game, first by releasing its own Moblin OS for 24 Page 5 of 11 Software Platforms netbooks in 200924 and then by merging it into MeeGo OS, a joint effort with Nokia, intended to run on netbooks, smartphones, and other computerbased devices.25 Of course, Intel's microprocessors are in much less danger of commoditization than hardware produced by PC and mobile handset OEMs. Nevertheless, despite its dominance of the microprocessor market for PCs, Intel's fate has oftentimes been too dependent on Microsoft, which has generally had the upper hand in the PC partnership and has been able to extract a larger share of overall profits.26 This explains Intel's efforts to play a significant role in the operating system layer—in PCs as well as new markets such as netbooks and smartphones. (p. 72) Although both mechanisms of across-layer SP competition have become particularly salient during the past four years, the second one is on the verge of inducing significantly more profound changes in the structure of computer-based industries. This is why it constitutes the focus of this section. Indeed, the emergence of new software platforms above (cloud-based platforms) and below (virtualization) the operating system are increasingly challenging its traditional status as the key MSSP in computer-based industries. As these alternative software platforms expand the range of application programming interfaces (APIs) they are able to provide to application developers, they diminish the economic value that is created and appropriated by the OS. 4.1. Cloud-Based Software platforms The tremendous potential of web-based platforms was largely anticipated in chapter 12 in IE through two case studies—eBay and Google. But while IE emphasized the disruptive effects of web-based software platforms on traditional industries such as retailing and newspapers (the disruption continues unabated), the focus here is on the way in which these SPs are diverting attention and value away from operating systems. To some extent, the advent of cloud-based SPs is the second-coming of the middleware visions associated with Netscape's Navigator browser and Sun's Java software platform in the late 1990s. Both Netscape Navigator and Java had been perceived by some, at the time, as serious threats to the lock held by Microsoft's Windows SP on third-party developers of PC applications and users.27 The vision was that they would become ubiquitous SPs, sitting on top of any OS and exposing their own APIs, so that third-party application developers would no longer have to worry about the underlying OS and hardware: they could write an application once and it would run “everywhere” (that is, on any computing environment supported by Netscape and/or Java). In other words, had they been successful, Netscape and Java would have significantly weakened the indirect network effects surrounding the Windows operating system. That vision did not quite materialize. Netscape's browser was completely marginalized by Microsoft's Internet Explorer during the first “browser wars” of the late 1990s (cf. Yoffie (2001)). Java did succeed in attracting a significant following of developers of “applets,” that is, small, web-based applications, but never seriously challenged the Windows SP for more sophisticated PC applications. Indeed, given the state of PC computing power and Internet connection speeds at the time, the scope for web-based applications was quite limited relative to applications running directly on the OS. Today, more than ten years later, things have radically changed. The uninterrupted progression of Moore's Law (the number of transistors which can be placed on an integrated circuit doubles roughly every two years) and the ubiquity of fast broadband connectivity have largely eliminated the limitations imposed on web-based (or cloud-based) applications. The last four years in particular have (p. 73) witnessed an explosion in the variety of computer-based, Internet-enabled devices beyond PCs, such as smartphones, netbooks and tablets. This in turn has generated increasing user demand for accessing the same applications and content across all of these devices. These developments have brought cloud computing and cloud applications into the mainstream. Put simply, “cloud computing” refers to a model of delivery of software applications and content in which they are hosted remotely by their providers on their own servers and accessed by users through an Internet connection and a browser (as opposed to residing on each individual user's device). Cloud applications range from consumer applications such as social networking (e.g., Facebook), online games (e.g., World of Warcraft), search (e.g., Bing), office productivity (e.g., Google docs) and personal finance (e.g., Mint), to enterprise applications such as customer relationship management (e.g., Salesforce's CRM). The development of cloud-based applications has naturally been accompanied by the development of cloud-based software platforms, which are a modern and more sophisticated version of the middleware software platforms of the late 1990s. Some of the most prominent cloud-based SPs that have emerged during the past four years are Amazon's Web Services, Facebook's Platform, Google's App Engine and various collections of APIs, Salesforce's Force.com. These four software platforms illustrate the wide variance in the nature and scope of services cloud-based SPs offer to third-party developers. At one end of the spectrum, the core of Amazon's Web Services (AWS) consists of “infrastructure services” such as data storage and computing power.28 AWS do also include a few APIs, mostly targeted at e-commerce applications (e.g., an API which enables websites affiliated with Amazon.com to easily draw and publish information about Amazon products on their own online properties). Still, the two most widely used AWS are Amazon Elastic Compute Cloud (EC2) and Amazon Simple Storage Service (S3) (Huckman et al. (2008)). These two services allow third-party developers to rent computing power and storage capacity on Amazon's servers, while paying based on usage. Google's App Engine is similar to and competes with AWS: it offers a similar range of infrastructure services, although with a stronger focus on web applications.29 But App Engine is surrounded by a broader set of APIs which Google has released over the years (e.g., Maps API, AdSense API, AdWords API). Thus, while both App Engine and AWS are considered to be “Platforms as a Service” (PaaS), App Engine is closer in nature to a true SP, that is, an operating system for web applications. At the other end of the spectrum, Facebook's Platform is purely a SP. Launched in 2007, it consists exclusively of APIs and programming tools allowing third-party developers to build applications for the social network (Eisenmann et al. (2008)). It does not offer any infrastructure services similar to AWS or App Engine. Salesforce's Force.com is somewhere in-between. It is a cloud-based platform providing both APIs and hosting services to third-party developers of cloud-based business applications. Second, as mentioned earlier, cloud-based SPs are not necessarily multisided, even when adopted by third-party application developers. Facebook's Platform is (p. 74) clearly multisided.30 The applications built on it are intended for and only available to Facebook users.31 The more such applications become available, the more valuable Facebook's network is for users and vice versa—the attractiveness of Facebook's Platform to developers is directly increasing in the number of Facebook users. On the other hand, Amazon's Web Services are mostly one-sided: the large majority of the third-party developers who use them are not affiliated with Amazon's e-commerce platform. Their adoption and usage of AWS does not increase the value of Amazon.com to its users or merchants. The only exception is Amazon's e-commerce API, which can be used to build functionalities for Amazon-affiliated merchants. But it is also used by a large number of independent e-commerce sites, which have no relationship to Amazon.com. Google's App Engine is similar to Amazon's Web Services, in that it targets generic websites and application developers, which a priori do not generate indirect network effects around Google's search engine. But some of Google's APIs (e.g., AdWords) clearly enhance indirect network effects around its search engine. Page 6 of 11 Software Platforms But if cloud-based SPs can be one-sided, then how does the other side—end-users—get access to the applications built on top of these SPs? The short answer is: through web browsers. It should not be surprising then that the coming of age of cloud computing during the past four years coincides with the advent of the second Browser Wars. The First Browser Wars—between Microsoft and Netscape—ended around the year 2000, when Internet Explorer (IE) had achieved more than 90 percent market share and Navigator had all but disappeared. After several years of almost uncontested dominance, IE found itself seriously challenged for the first time by Mozilla's Firefox. Launched in November 2004, Firefox reached over 10 percent market share within a year and 23 percent by the summer of 2010.32 Until 2008, Firefox accounted for most of Microsoft's Internet Explorer's decline, which went from more than 85 percent market share in 2004 to less than 60 percent in February 2011. But the new Browser Wars truly heated up with the launch of Google's Chrome browser in September 2008. Within 3 years, Chrome reached more than 15 percent market share. Around the same time, Apple's Safari browser also started to pick up serious steam, largely due to the popularity of the iPhone: it passed the 10 percent market share mark for the first time in its existence sometime during 2011.33 And Microsoft itself has intensified its efforts with IE–a significantly improved IE 9 was launched in March 2011.34 The key factor driving the Second Browser Wars is the crucial status of browsers as users’ gateway to web-based applications. Although browsers have to rely on open web standards (which by definition are available to and on any browser), they can still exert significant influence on how these applications are being used and ultimately monetized. For companies like Apple, Google and Microsoft, browsers can be leveraged as indirect strategic weapons to drive up the value of other, complementary assets.35 Take Google for instance. The first and most immediate benefit of its Chrome browser is that it can be used to increase traffic to Google's core asset, the search engine, by making it the default home page (many users rarely (p. 75) modify the default). It also enables Google to obtain more information on user browsing behavior, which allows it to further improve its mechanism for serving search-related advertising. Finally, provided Chrome achieves significant market share, it enables Google to influence the evolution of webbased APIs, which can be used to increase the value of Google's cloud-based SPs (App Engine and its collections of APIs). In this context, it is not surprising that Google took the logical next step in 2009, by announcing a Chrome operating system based on the Chrome browser and the other four sets of Google APIs.36 Chrome OS will exclusively support web applications, that is, applications that are accessible through the Internet and are not tied to a specific device.37 Regardless of whether they use one-sided or multisided business models and regardless of their scope and nature (that is, whether they offer software building blocks only or also include infrastructure services), cloud-based SPs are drawing an increasing number of developers. Their expanding scope and functionalities make it more attractive (from a business standpoint) to write web-based applications as opposed to operating system-based applications. Of course, this transition is unlikely to go all the way to its logical extreme, in which all applications would be written for cloud-based SPs and would no longer be tied to a specific operating system's APIs. Indeed, while cloud-based applications have many economic advantages (consistent and easy use through many devices; easy installation and upgrades), there are some important tradeoffs. Users (particularly corporate) may worry about security and having control over their data, which implies a preference for having at least parts of the applications reside on their local computers or other devices. Furthermore, the cross-device and cross-operating system nature of cloud applications implies that lowest-common denominator compromises must sometimes be made in building functionalities. This can result in disadvantages relative to applications which reside on local devices and can therefore be optimized for specific operating systems or hardware. Finally, in the enterprise software market, there is a large ecosystem of integration and services providers, which have vested financial incentives to resist the shift toward purely cloud-based applications—the latter would reduce the need for the existence of these companies. Consequently, although the “write once, run everywhere” promise of cloud-based SPs is very attractive for many application developers, there is a significant proportion of applications (particularly for enterprise customers) which need to rely on operating system APIs, so that the OS's power as a SP is unlikely to be entirely supplanted by cloud-based SPs. The rate of adoption of the four cloud-based platforms mentioned above has been staggering: Force.com has over 1000 applications; 38 more than 300,000 developers are active on App Engine and AWS; 39 and Facebook Platform supports over 550,000 applications.40 But these numbers should not be taken at face value. The majority of applications built on these SPs are still small-scale web applications; furthermore, Facebook's applications have oftentimes been characterized as useless time-wasters.41 Despite all these limitations however, the trend is undeniable: the battle for creating innovative applications—along with (p. 76) the value of the corresponding network effects— is moving away from the operating system and onto the web. 4.2. Virtualization (Note: This subsection is primarily based on Yoffie, Hagiu, and Slind, 2009.) While cloud-based (multisided) software platforms tend to move economic value upward from the operating system, virtualization arguably moves value in the opposite direction. Up until the early 2000s, it was hard if not impossible to imagine a world in which the operating system would not be the first software layer on top of a computer device's microprocessor, or in which the relationship between the microprocessor and the operating system would not be one-to-one. Virtualization has challenged both of these assumptions. Virtualization is a software technology that was first developed in the 1960s by IBM in order to improve the utilization of large mainframe computers. It did so by partitioning the underlying mainframe hardware and making it available as separate “virtual machines,” to be utilized simultaneously by different tasks. The modern version of virtualization was invented and commercialized by VMware Inc. in the early 2000s and applied to the x86 family of servers and personal computers. In its most popular incarnation (called hypervisor or “bare-metal” approach), the virtualization “layer” sits between the microprocessor and the operating system. As illustrated in the Figure 3.1, it creates replicas of the underlying hardware, which can then be used as virtual machines running independently of one another and with separate operating systems. Figure 3.1 Virtualization Technology in the Hypervisor Mode. Page 7 of 11 Software Platforms The most immediate benefit of virtualization was to enable enterprises to consolidate underutilized servers, thereby achieving large technology cost reductions. Today, there are three additional components to the economic value created by virtualization. First, the virtual machines are isolated from one another, which is highly desirable for reasons related to security (one virtual machine crashing has no effect on the others) and flexibility (the virtual machines can be handled independently). Second, the virtual machines encapsulate entire systems (hardware configuration, operating system, applications) in files, which means that they can be easily transferred and transported, just like any other files (e.g., on USB keys). (p. 77) Third, the virtual machines are independent of the underlying physical hardware, so they can be moved around different servers. This is particularly useful when one physical server crashes or when one desires to economize energy by shutting down some servers at off-peak times. The tremendous value of virtualization is reflected in the quick rise of VMware, the market leader. VMware was founded in 1998 by five Stanford and Berkeley engineers and during the next 10 years became one of the hottest companies in the technology sector. After a $20 million funding round in May of 2000, the company drew acquisition interest from Microsoft in 2002. But Microsoft eventually backed off in the face of what it perceived to be an excessive asking price. VMware was then acquired in 2003 for $635 million by the storage system giant EMC. In 2007, EMC sold 10 percent of VMware shares in an IPO, which earned it $957 million at the initial offering price of $29 a share. As of October 2010, VMware's share price was around $80, implying a market capitalization of $33 billion. While a significant portion of this value reflects the large cost savings enabled by virtualization, a key reason for the hype surrounding VMware is the disruptive potential of its software platform for the established order in the server and PC software industries. The basic virtualization technology—the hypervisor—is a SP because it exposes APIs which can be utilized by applications running on top of it. It is in fact a two-sided software platform: there are hundreds of independent software vendors developing applications for VMware's vSphere hypervisor (VMware also provides many applications of its own, just like Microsoft does for Windows). Examples of virtualization applications are those that enable corporate IT departments to move virtual machines across physical servers without interruption (the corresponding application from VMware is called VMotion), to test and deploy new services, and so forth. There are three important ways in which virtualization threatens to undermine the power held by the operating system (Windows in particular) over the other layers in the server and PC stacks. First, by breaking the one-to-one relationship between hardware and operating system, virtualization makes it much easier for individual users and companies to use multiple OSs. In other words, thanks to virtualization, the user side of the operating system MSSP becomes a multihoming side instead of a single-homing side. Thus, while virtualization may not necessarily decrease the absolute number of Windows licenses sold, it does increase the market penetration of alternative operating systems (e.g., Linux for enterprises, Mac OS for consumers). In turn, this increases application developers’ incentives to (also) develop for those other operating systems: if a given user does not have the required operating system for running the developer's application to begin with, she can use virtualization to install it alongside Windows, without having to buy a new machine. Overall, this mechanism significantly weakens the famous “applications barrier to entry” enjoyed by Windows. Second, because it is a software platform sitting at the same level or below the OS, virtualization inevitably takes away some of the OS's value by offering APIs that otherwise might have been offered by the OS. While this effect is limited today given that the applications built on top of a hypervisor are quite different in nature (p. 78) from the ones built on top of an OS, it will likely grow in importance as virtualization technology (VMware in particular) expands its scope. Third, virtualization enables the creation of “virtual appliances,” that is, applications that come bundled with their own, specialized OS. Instead of building on top of large, all-purpose OSs like Windows, a developer can match her application with a stripped-down OS that she chooses according to criteria including security, cost, and programming ease. She can then distribute the OS-application bundle, knowing that it will run on any computing system equipped with the virtualization platform. It is clear then that, if widely adopted, the virtual appliance model would have the effect of commoditizing the OS. By the end of 2009, VMware's ISV partners had brought more than 1,300 virtual appliances to market.42 These developments have of course not been lost on Microsoft. In response to the threats posed by VMware's virtualization technology, Microsoft launched its own virtualization platform in 2008. Called Hyper-V, it was bundled with the Windows Server OS, thus being effectively priced at 0 and benefitting from the powerful Windows distribution channels. But the fact that Microsoft designed Hyper-V to work closely with Windows can also be perceived as a weakness relative to VMware's OS-agnostic approach. The battle between the two companies will be a fascinating one to follow over the next few years. To some extent, it is reminiscent of the browser wars between Microsoft and Netscape during the late 1990s. Both involve highly innovative companies with superior technologies, which pose disruptive threats to Microsoft's Windows software platform. Indeed, VMware today has a significant technical and first-mover advantage, which it is seeking to maintain by building high-value applications on top of its virtualization platform.43 These applications would increase the switching costs for the customers who are already using VMware technology and make them less likely to switch to Microsoft's competing Hyper-V software platform. It is the nature of those customers and of the technology involved that makes the comparison to the browser wars somewhat limited. Large corporations (who make up the large majority of VMware's customers) adopting a virtualization platform for their data centers are much less easily swayed than consumers choosing browsers. But regardless of the duration and the outcome of the virtualization platform wars, what is certain is that the appearance of virtualization as a new SP is irreversibly changing the way in which economic value is created and extracted in the relevant technology stacks. 5. Conclusion In this chapter I have attempted to briefly survey some of the developments related to software platforms since the publication of IE in 2006, which I believe are most interesting for economics and management researchers. (p. 79) Whereas IE had emphasized the inherently multisided nature of SPs, the diverse nature of cloud-based SPs suggests that this view might be too narrow. By definition, the user side of web-based SP is always on board as long as it has access to an Internet-connected device and a browser. In this context, multi-sidedness becomes a subtler issue: it essentially depends on whether the provider of the SP owns a related product or service, whose value is increased by the applications built upon the SP. Furthermore, while multi-sidedness holds the potential of creating larger value through indirect network effects, it might also create conflicts of interest with some of the members of the relevant ecosystem. This suggests that some software platforms may face interesting strategic tradeoffs when choosing to function as one-sided or as multisided businesses. Regarding SP business models, the battle between Microsoft's three-sided SP (Windows) and Apple's two-sided SP (the Macintosh computer) from the PC market is currently being replicated by Google and Apple in the smartphone industry—quite likely with different results. Comparing and contrasting Page 8 of 11 Software Platforms these two battles should provide substantial research material for analyzing the impact of industry characteristics on the outcome of competitive dynamics between platforms with different degrees of vertical integration. The evolution of SP pricing structures, with more and more SPs (aside from videogame console makers) charging the application developer side, raises the intriguing question of whether we might be witnessing a shift toward more balanced pricing models, in which both sides (developers and users) directly contribute substantial shares of platform profits. Governance rules also seem to have emerged as an important strategic instrument for SPs. A promising avenue for future research would be to investigate the conditions under which governance rules can provide a sustainable source of differentiation between otherwise similar SPs competing in the same industry. Finally, the competitive dynamics between software platforms operating at different levels of the technology stack (e.g., web-based platforms, operating systems, virtualization platform) points to important research questions, which, to the best of my knowledge, have not been systematically explored. In particular, is it possible to predict, solely based on exogenously given industry characteristics, the layer in the relevant “stack” at which the most valuable and powerful (multisided) platform should arise? And if not, that is, if multiple outcomes are possible depending on the actions chosen by players in different layers, can one characterize the general strategies that enable players in one layer to “commoditize” the other layers? Acknowledgment I am very grateful to my two co-authors on Invisible Engines—David Evans and Richard Schmalensee—for helpful discussions and feedback on an early draft. The views expressed in this chapter (and any related misconceptions) are, however, entirely my own. References Boudreau, K., 2008. Too Many Complementors? London Business School, working paper. Boudreau, K., Hagiu, A., 2009. Platform Rules: Multi-Sided Platforms as Regulators. In: Gawer, A. (Ed.), Platforms, Markets and Innovation, Edward Elgar, Cheltenham and Northampton, pp. 163–191. Casadesus-Masanell, R., Yoffie, D.B., 2007. Wintel: Cooperation and Conflict. Management Science 53, pp. 584–598. Eisenmann, T., Piskorski, M.J., Feinstein, B., Chen, D., 2008. Facebook's Platforms. Harvard Business School Case No. 808–128. Evans, D. S., Hagiu, A., Schmalensee, R., 2006. Invisible Engines: How Software Platforms Drive Innovation and Transform Industries. The MIT Press. Hagiu, A., 2009. Note on Multi-Sided Platforms—Economic Foundations and Strategy. Harvard Business School Note No. 709–484. Hagiu, A, Halaburda, H., 2009. Responding to the Wii. Harvard Business School Case No. 709–448 and Teaching Note No. 709–481. Hagiu, A., Yoffie, D.B., 2009. What's Your Google Strategy? Harvard Business Review April Issue, 1–9. Hagiu, A., Yoffie, D.B., Slind, M., 2007. Brightcove and the Future of Internet Television. Harvard Business School Case No. 707–457 and Teaching Note No. 707–568. Huckman, R. S., Pisano, G., Kind, L., 2008. Amazon Web Services. Harvard Business School Case No. 609–048. Yoffie, D.B., 2001. The Browser Wars: 1994–1998. Harvard Business School Case No. 798–094. Yoffie, D.B., 2003a. Wintel (B): From NSP to MMX. Harvard Business School Case No. 704–420. Yoffie, D.B., 2003b. Wintel (C): From MMX to the Internet. Harvard Business School Case No. 704–421. Yoffie, D.B., Casadesus-Masanell, R., Mattu, S., 2004. Wintel (A): Cooperation or Conflict? Harvard Business School Case No. 704–419. Yoffie, D.B., Hagiu, A., Slind, M., 2009. VMware, Inc., 2008. Harvard Business School Case No. 709–435. Notes: (1.) See Goyal, S. “Social Networks on the Web,” chapter 16 in this book. (2.) “Lexmark Tries to Catch App Fever,” Wall Street Journal, October 26th 2010. (http://online.wsj.com/article/SB10001424052702303467004575574333014975588.html) (3.) For related discussions in this book, see Lee, R. “Home Videogame Platforms,” chapter 4; Jullien, B. “Two-Sided B2B Platforms,” chapter 7; Choi, J.P. “Bundling Information Goods,” chapter 11; Anderson, S. “Advertising on the Internet,” chapter 14. (4.) Chapter 2 in IE provides a detailed technical and economic background on software platforms. (5.) Today Facebook is a four-sided platform: it connects (1) users; (2) third-party application developers; (3) advertisers; (4) third-party websites that can be browsed with one's Facebook identity via Facebook Connect. (6.) The term “platform governance” was coined by Boudreau and Hagiu (2009). (7.) This side includes hardware OEMs as well as suppliers of accessories such as jogging sleeves for iPods or iPhones. (8.) http://www.gartner.com/it/page.jsp?id=1466313 (9.) “RIM's Blackberry: failure to communicate,” Businessweek, October 7th 2010. (http://www.businessweek.com/magazine/content/10_42/b4199076785733.htm) (10.) “Samsung tries to woo TV app developers,” CNET, August 31st 2010. (http://news.cnet.com/8301–31021_3–20015215-260.html#ixzz124u0TYXs) Page 9 of 11 Software Platforms (11.) For more on digital formats and piracy, see Belleflamme, P. and M. Peitz “Digital Piracy: Theory,” chapter 18 and Waldfogel, J. “Digital Piracy: Empirics,” chapter 19. (12.) See Lee, R. “Home Videogame Platforms,” chapter 4 (13.) “Amazon Amps Up Apps Rivalry,” Wall Street Journal, October 7th 2010. (http://online.wsj.com/article/SB10001424052748704696304575538273116222304.html) (14.) A well-known example is Apple's refusal to allow applications relying on Adobe's Flash program. In the face of mounting developer criticism and the risk of antitrust investigation, Apple recently decided to relax some of its restrictions. See: “Apple Blinks in Apps Fights,” Wall Street Journal, September 10th 2010. (http://online.wsj.com/article/SB10001424052748704644404575481471217581344.html) (15.) “Apple's iAds Policy Gets FTC Scrutiny,” International Business Times, June 14th 2010. (http://www.ibtimes.com/articles/28542/20100614/apples-iad-policy-comes-under-ftc-scanner.htm) (16.) The only other software platform studied in IE that charged third-party developers was NTT DoCoMo's i-mode (9 percent cut of revenues for content providers choosing to rely on i-mode's billing system). But even in that case, the revenues coming from developers were less than 1 percent of total i-mode revenues. (17.) http://ftalphaville.ft.com/blog/2010/07/13/285006/goldman-really-likes-its-new-ipad/ (18.) http://gigaom.com/2010/01/12/the-apple-app-store-economy/ (19.) The 40 minutes per day estimate is based on Facebook's announcement that its more than 500 million users spend over 700 billion minutes on the site each month. See: http://www.facebook.com/press/info.php?statistics (accessed December 6th 2010). (20.) Some aspects of the competitive dynamics among complementors are explored by Casadesus-Masanell and Yoffie (2007). (21.) Chapter 7 in IE provides background on the founding of Symbian and its evolution up to 2005. (22.) “Samsung Gives Details on Bada OS,” InformationWeek, December 9th 2009. (http://www.informationweek.com/news/mobility/business/showArticle.jhtml?articleID=222001332) (23.) “Samsung tries to woo TV app developers,” CNET, August 31st 2010. (http://news.cnet.com/8301–31021_3–20015215-260.html#ixzz124u0TYXs) (24.) “Intel Adopts an Identity in Software,” New York Times, May 25th 2009. (http://www.nytimes.com/2009/05/25/technology/businesscomputing/25soft.html) (25.) “Intel and Nokia Team Up on Mobile Software,” New York Times, February 15th 2010. (http://bits.blogs.nytimes.com/2010/02/15/intel-and-nokiateam-up-on-mobile-software/) (26.) For details on the Microsoft-Intel relationship, see Yoffie (2003a) and (2003b), Yoffie et al. (2004) and CITE Wintel (A) case and “The End of Wintel,” The Economist, July 29th 2010 (http://www.economist.com/node/16693547). (27.) U.S. v. Microsoft—Proposed Findings of Fact (http://www.justice.gov/atr/cases/f2600/2613pdf.htm). (28.) http://aws.amazon.com/ (29.) “Comparing Amazon's and Google's Platform-as-a-Service (PaaS) Offerings,” ZDNet, April 11th 2008. (http://www.zdnet.com/blog/hinchcliffe/comparing-amazons-and-googles-platform-as-a-service-paas-offerings/166) (30.) Facebook is currently a four sided platform. It connects: (1) users; (2) third-party application and content developers; (3) advertisers; and (4) third-party web properties that can be browsed using one's Facebook identity via Facebook Connect. (31.) This is similar to the eBay SP described in chapter 12 in IE: eBay's APIs are specifically designed for building applications targeted at eBay buyers and sellers. (32.) http://www.netmarketshare.com/browser-market-share.aspx?qprid=0, accessed March 22, 2011. (33.) “Microsoft Faces New Browser Foe in Google,” The New York Times, September 1st 2008. (http://www.nytimes.com/2008/09/02/technology/02google.html) “Google Rekindles Browser Wars,” Wall Street Journal, July 7th 2010. (http://online.wsj.com/article/SB10001424052748704178004575351290753354382.html) (34.) “Microsoft Modernizes Web Ambitions with IE9,” CNET News, March 16th 2010. (http://news.cnet.com/8301-30685_3-20000433-264.html) (35.) “Microsoft Faces New Browser Foe in Google,” The New York Times, September 1st 2008. (http://www.nytimes.com/2008/09/02/technology/02google.html) “Google Rekindles Browser Wars,” Wall Street Journal, July 7th 2010. (http://online.wsj.com/article/SB10001424052748704178004575351290753354382.html) (36.) “Clash of the Titans,” The Economist, July 8th 2009. (http://www.economist.com/node/13982647? story_id=13982647&source=features_box_main) (37.) “Google Announces Chrome OS,” PCWorld, July 8th 2009. (http://www.pcworld.com/article/168028/google_announces_chrome_os.html) (38.) http://sites.force.com/appexchange/browse?type=Apps (39.) “Who's Who in Application Platforms for Cloud Computing: The Cloud Specialists,” Gartner Report, September 2009. http://en.wikipedia.org/wiki/Amazon_Web_Services (40.) http://www.facebook.com/press/info.php?statistics Page 10 of 11 Software Platforms (41.) “Useless Applications Plague Facebook,” The Lantern, June 20th 2009. (http://www.thelantern.com/2.1345/useless-applications-plague-facebook1.76407) (42.) VMware 2010 10-K filing. Available at: http://ir.vmware.com/phoenix.zhtml?c=193221&p=IROLsecToc&TOC=aHR0cDovL2lyLmludC53ZXN0bGF3YnVzaW5lc3MuY29tL2RvY3VtZW50L3YxLzAwMDExOTMxMjUtMTAtMDQ0Njc3L3RvYy9wYWdl&ListAll=1 (43.) “VMware Lays Down Corporate IT Marker,” Financial Times, October 4, 2010. Andrei Hagiu Andrei Hagiu is Assistant Professor of Strategy at the Harvard Business School. Page 11 of 11 Home Videogame Platforms Oxford Handbooks Online Home Videogame Platforms Robin S. Lee The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0004 Abstract and Keywords This article highlights the key features that distinguish the industrial organization of videogames from other similar hardware-software markets. The models of consumer and firm (both hardware and software) strategic behavior are then addressed. Videogame consoles sit squarely amidst the convergence battle between personal computers and other consumer electronic devices. Games developed for one console are not compatible with others. The pricing of hardware consoles has been the most developed area of research on platform strategy. It is noted that even retrospective analysis of the videogame market when all the players were known needed sophisticated modeling techniques to handle the industry's complexities, which include controlling for dynamic issues and accounting for consumer and software heterogeneity. The success and proliferation of videogames will continue to spawn broader questions and enhance the understanding of general networked industries. Keywords: videogame industry, hardware, software, videogame consoles, pricing, videogame market, consumer 1. Introduction What began with a box called Pong that bounced a white dot back-and-forth between two “paddles” on a television screen has now blossomed into a $60B industry worldwide, generating $20B annually in the United States alone.1 Today, videogames are a serious business, with nearly three-quarters of US households owning an electronic device specifically used for gaming, and many predicting that figure to increase in the coming years.2 Given the widespread adoption of a new generation of videogame systems introduced in 2006 and the ever growing popularity of online and on-the-go gaming, videogames are also no longer strictly the stuff of child's play: surveys indicate 69 percent of US heads of households engage in computer and videogames, with the average age of a player being 34 years old.3 As newer devices continue to emerge with even more advanced and immersive technologies, it is likely that videogames will continue to play an ever increasing role in culture, media, and entertainment. Owing no small part to this success, the videogame industry has been the subject of a growing number of studies and papers. This chapter focuses on research within a particular slice of the entire industry – the home videogame console market – which on its own is a fertile subject for economic research, both in theory and empirics. As a canonical hardware-software market rife with network effects (c.f., Katz and Shapiro, (1985); Farrell and Saloner (1986)), videogames are an ideal setting to apply theoretical models of platform competition and “two-sided markets”; and as a vertical market dominated on different sides by a small number of (p. 84) oligopolistic firms, videogames provide an opportunity to study issues related to bilateral oligopoly and vertical contracting. Furthermore, with the development of new empirical methods and tools, data from the videogame market can be used to estimate sophisticated models of dynamic consumer demand, durable goods pricing, and product investment and creation. By focusing solely on videogames, economic research can inform our analysis of other Page 1 of 18 Home Videogame Platforms related markets in technology, media, or even more broadly defined platform-intermediated markets. This chapter is organized as follows. I first provide a brief overview of the industrial organization of videogames, and emphasize the key features that distinguish it from other similar hardware-software markets. Second, I survey economic research on videogames, focusing primarily on models of consumer and firm (both hardware and software) strategic behavior; I also highlight potential avenues for future research, particularly with respect to platform competition. Finally, I conclude by discussing how these economic models can help us better understand vertical and organizational issues within the industry, such as the impact of exclusive contracting and integration between hardware platforms and software developers on industry structure and welfare. 2. The Industrial Organization of Home Videogames 2.1. Hardware Today firms in a variety of industries produce hardware devices that vary widely in size, portability, and functionality for the purpose of electronic gaming. However, as has been the case for most of the four-decade history of the home videogame industry, these devices are primarily stationary “boxes” that require a monitor or television set for use. Referred to as consoles or platforms, these devices are standardized computers tailored for gaming and produced by a single firm. Approximately 53 percent of households in the United States are estimated to own a videogame console or handheld system.4 For the past decade, the three main console manufacturers or platform providers have been Nintendo, Sony, and Microsoft. Nintendo, originally a Japanese playing card company founded in the late 19th century, is the most experienced veteran of the three: it has manufactured videogame consoles since the late 1970s, and is the only firm whose primary (and only) business is videogames. Its Nintendo Entertainment System (NES), released first in Japan in 1983 and two years later in the United States, was the first major videogame platform to achieve global success. Nintendo has since released several consoles, including the most recent “Wii” in 2006; the Wii was one of the first to incorporate a novel motion-sensing interface and credited for expanding the appeal of home videogaming to a broader audience. (p. 85) The other two console manufacturers entered many years after Nintendo's NES system dominated the market. Sony released its first videogame console – the Playstation – in 1995.5 One of the first consoles to use games produced on CDs as opposed to more expensive cartridges, the Playstation would sell over 100M units in its lifetime and establish Sony as the dominant console manufacturer at the turn of the 21st century. The Playstation has had two successors: the PS2, released in 2000, became the best selling console in history with over 140M units sold; 6 and the PS3, released in 2006, is perhaps most famous for being one of the most expensive videogame consoles ever produced. Finally, Microsoft, as the newest of the three console manufacturers, entered the home videogame market in 2001 with the Xbox console; it later followed it up with the Xbox360 in 2005. Forty-one percent of US households are estimated to own at least one of the three newest consoles.7 In general, hardware specifications for a given console remain fixed over its lifetime to ensure compatibility with any games produced for that console; only by releasing a new console could a firm traditionally introduce new hardware with more powerful processing power and graphical capabilities. A new set of consoles has historically been launched approximately every five years – thus heralding a new “generation” within the industry; however, due to the large sunk-cost associated with development of new consoles, the desire of hardware manufacturers to recoup initial investments, and the shift toward upgrading existing consoles via add-on accessories, the length between new generations is likely to increase in the future.8 Although this chapter focuses on home videogame consoles, there is still a large market for dedicated portable gaming devices, currently dominated by Sony and Nintendo, and gaming on multifunction devices, such as smartphones and media players (e.g, Apple's iPod and iPhone). In addition, although personal computers (PCs) have always been able to play games, their significance as traditional videogame platforms is small: less than 5 percent of videogame software revenues today derive from PC game sales (though this may change in the future given the rise of online gaming via virtual worlds or social networks).9 Finally, just as other devices in other industries have been adding video-gaming capabilities, videogame consoles, Page 2 of 18 Home Videogame Platforms too, have been adding greater functionality: for example, today's consoles also function as fully independent media hubs with the ability to download and stream movies, music, television, and other forms of digital content over the Internet. Videogame consoles thus sit squarely amidst the convergence battle between personal computers and other consumer electronic devices. 2.2. Software and Games In addition to console manufacturers, the videogame industry also comprises firms involved in the production of software or games.10 These firms can be roughly categorized into two types: developers or development studios, who undertake (p. 86) the programming and creative execution of a game; and publishers, who handle advertising, marketing, and distribution efforts. This distinction is not necessarily sharp: most publishers are integrated into software development, often owning at least one studio; and although independent software developers exist, they often have close relationships with a single software publisher for financing in exchange for distribution and publishing rights. Such relationships may appear to be similar to integration insofar they are often exclusive, and have become standard for independent developers as the costs of creating games have dramatically increased over time.11 Console manufacturers also historically have been and continue to be integrated into software development and publishing. Any title produced by a console manufacture's own studios or distributed by its own publisher is referred to as a first-party title, and is exclusive to that hardware platform. All other games are third-party titles and are developed and published by other firms. Much like videogame hardware, videogame software is predominantly produced by a handful of large firms: the top 10 publishers, which also includes the main three console manufacturers, produce over 70 percent of all games sold, with the largest (Electronic Arts) commanding a 20 percent market share. Furthermore, individual games have been increasingly exhibiting high degrees of sales concentration with the emergence of “killer applications” and “hit games.” During the “sixth generation” of the industry between 2000 and 2005, nearly 1,600 unique software titles were released for the three main consoles; however, the top 25 titles on each system comprised 25 percent of total software sales, and the top 100 titles over 50 percent. Since then, a handful of titles have sold millions of copies, with some games even generating over $1B in sales on their own.12 Finally, unlike hardware, the lifetime of a particular game is fairly short: typically half of a game's lifetime sales occur within the first 3 months of release, and very rarely do games continue to sell well more than half a year from release. 2.3. Network Effects and Pricing Since consoles have little if any stand-alone value, consumers typically purchase them only if there are desirable software titles available. At the same time, software publishers release titles for consoles that either have or are expected to have a large installed base of users. These network effects operative on both sides of the market are manifest in most hardware-software industries, and are partly a reason for the complex form of platform pricing exhibited by videogame platforms: most platform providers subsidize the sale of hardware to consumers, selling them close to or below cost, while charging publishers and developers a royalty for every game sold (Hagiu, 2006; Evans, Hagiu, and Schmalensee, 2006). This “razor blade” model was initially used by Atari with the release of its VCS console in 1977 – Atari originally sold its hardware at a very slight margin, but its own (p. 87) videogame cartridges at a 200 percent markup (Kent, 2001) – and Nintendo was the first to charge software royalties with its NES system nearly a decade later.13 As a result, most platform profits have been and continue to be primarily derived not from hardware sales, but rather from software sales and royalties.14 Note this stands in contrast to the traditional pricing model in PCs, where the operating system (e.g., Microsoft Windows) is typically sold to the consumer at a positive markup, yet no royalties or charges are levied on third-party software developers. For the past two generations of videogame consoles, initial losses incurred by platform providers due to this pricing scheme have been substantial: e.g., the original Xbox had estimated production costs of at least $375 yet sold for an introductory price of $249; in the first four years of existence, Microsoft was estimated to have lost $4B on the Xbox division alone.15 However, as costs for console production typically fall over time faster than the retail price, a console manufacturer's profits on hardware typically increases in the later years of a generation: e.g., Sony in 16 Page 3 of 18 Home Videogame Platforms 2010 announced it finally was making a profit on its PS3 console 3.5 years after it launched; 16 a similar path to profitability was followed by Sony's PS2 after its release. There are, however, some exceptions: e.g., Nintendo's current generation Wii console was profitable from day one and was never sold below cost.17 2.4. Porting, Multi-homing, and Exclusivity Typically within a generation, games developed for one console are not compatible with others; in order to be played on another console, the game must explicitly be “ported” by a software developer and another version of the game created.18 Due to the additional development and programming time and expense to develop additional versions of the game, the porting costs of supporting an additional console can be as high as millions of dollars for the current generation of systems.19 During the early years of the videogame industry (pre-1983), any firm who wished to produce a videogame could develop and release a game for any console which utilized cartridges or interchangeable media. However, many console manufacturers recouped console development costs from the sale of their own games, and saw the sale of rival third-party games as a threat; some console manufacturers even sued rival software developers to (unsuccessfully) keep them off their systems (Kent, 2001). The inability to restrict the supply of third-party software led to a subsequent glut of games released in the early 1980s, many of low quality; in turn, this partially caused the videogame market crash of 1983 in which demand for videogame hardware and software suddenly dried up. Whereas there used to be over a hundred software developers in 1982, only a handful remained a year later. As one of the survivors of the crash, Nintendo deviated from the strategy employed by previous console manufacturers when releasing its NES console in the United States in 1985. First, it actively courted third-party software developers, understanding that a greater variety of software would increase attractiveness of the platform to other consumers; at the same time, it prevented unauthorized games (p. 88) from being released via a security system in which cartridges without Nintendo's proprietary chip could not be played on its console. Nintendo also imposed other restrictions on its third-party software licensees: each developer was limited to publishing only 5 games a year, had to give the NES exclusivity on all games for at least 2 years, and had to pay a 20 percent royalty on software sales (Kent, 2001; Evans, Hagiu, and Schmalensee, 2006). It was not until 1990 that Nintendo – in the midst of lawsuits levied by competitors and an FTC investigation for anticompetitive behavior – announced that it would no longer restrict the number of games its developers could produce or prohibit them from producing games for other systems.20 Since then, forced exclusivity – the requirement that a videogame be only provided for a given console or not at all – has not been used in the industry.21 Though many software titles now choose to “multihome” and support multiple consoles, there are still instances in which third-party games are exclusive: some do so voluntarily (perhaps due to high porting costs), some engage in an exclusive publishing agreement with the console provider (typically in exchange for a lump sum payment), and others may essentially integrate with a platform by selling the entire development studio outright. 2.5. Consumers As mentioned earlier in this chapter, the vast majority of videogame players are no longer children: 75 percent of gamers are 18 years old or older, with two-thirds of those between the ages of 18–49.22 In addition, there is a wide degree of variance in usage and purchasing behavior across consumers: in 2007, Nielsen estimated the heaviest using 20 percent of videogame players accounted for nearly 75 percent of total videogame console usage (by hours played), averaging 345 minutes per day. Furthermore, although on average 6–9 games were sold per console between 2000–2005, “heavy gamers” reported owning collections of over 50+ games,23 and on average purchased more than 1 game per month.24 3. Economics of the Videogame Industry Although there exist active markets for PC and portable gaming, most research on videogames has focused on the home videogame market. This is not without reason. First, the home videogame industry is convenient to study since all relevant firms within a generation are known, and there exist data containing a list of all software Page 4 of 18 Home Videogame Platforms produced for nearly all of the consoles released in the past three decades. Compare this to the PC industry, where there are thousands of hardware manufacturers and product varieties, and even greater numbers of software developers and products; obtaining detailed price and quantity information, for example, on (p. 89) the universe of PC configurations, accessories, and software products would be infeasible. Second, there are relatively few substitutes to a home videogame console, allowing for a convenient market definition. Finally, as videogame consoles have been refreshed over time, there is the potential for testing repeated market interactions across multiple generations. This section provides a brief (and thus by no means comprehensive) review of recent economic research on the home videogame industry, and emphasizes both the advances and limitations of the literature in capturing important features and dynamics of the market. As understanding the interactions between the three major types of players in the industry – software firms, hardware firms, and consumers – provide the foundation for any subsequent analysis (e.g., how industry structure or welfare changes following a policy intervention or merger), it is unsurprising that the vast majority of papers have first focused on modeling the strategic decisions of these agents. Only with these models and estimates in hand have researchers have begun addressing more complicated questions including the role and impact of vertical relationships in this industry. Across all of these fronts remain several open questions, and I highlight those areas in which future study would prove useful. 3.1. Consumer Demand and Software Supply As with many hardware-software industries, videogames exhibit network effects in that the value of purchasing a videogame console as a consumer increases in the number of other consumers who also decide to purchase that console. Although there is a direct effect in that people may prefer owning the same videogame console as their friends or neighbors, the primary means by which this occurs is through an indirect effect: more consumers onboard a particular console attracts more games to be produced for that console, which in turn makes the console an even more desirable product.25,26 Such indirect network effects also work in the other direction: software developers may benefit from other software developers supporting the same console in that more games attracts more consumers, which further increases the potential returns for developing a game for that console.27 Numerous studies have attempted to empirically document or measure the strength, persistence, and asymmetry of these kinds of network effects in a variety of industries. Many of these original empirical papers base their analysis on models in which consumers of competing platforms preferred to purchase the device with a greater number of compatible software titles. As long as consumers preferred a greater variety of software products – typically modeled via CES preferences – and certain assumptions on the supply of software held, then a simple loglinear relationship between consumer demand for hardware and the availability of software could be theoretically shown to arise in equilibrium (c.f Chou and Shy (1990); Church and Gandal (1992)). Empirical research based on these types of models include studies on adoption of CD players (Gandal, Kende, and Rob, 2000), DVD (p. 90) players (Dranove and Gandal, 2003), VCRs (Ohashi, 2003), and personal digital assistants (Nair, Chintagunta, and Dubé, 2004). In the spirit of this literature, Shankar and Bayus (2003) and Clements and Ohashi (2005) are two of the earliest papers to empirically estimate the existence and magnitude of network effects in the videogame industry. Whereas Shankar and Bayus (2003) assume software supply is exogenous and not directly affected by hardware demand, Clements and Ohashi (2005) estimate two simultaneous equations representing the two-sided relationship between a console's installed base of users and its games. This approach, motivated by a static model of consumer demand and software supply, is followed by other papers analyzing the videogame industry (e.g., Corts and Lederman (2009); Prieger and Hu (2010)), and is useful to describe briefly here. The model assumes a consumer's utility from purchasing console j at time t is given by: where xj are console j's observable characteristics (e.g., speed, processing power), pj,t the price, Nj,t the number of available software titles for console j, ξj,t an error unobservable to the econometrician, and εj,t a standard logit Page 5 of 18 Home Videogame Platforms error; a consumer purchases console j at time t if it delivers the maximum utility among all alternatives (including an outside option). As in Berry (1994), this can be converted into a linear regression via integrating out the logit errors and using differences in log observed shares for each product to obtain the following estimating equation: (1) where sj,t, so,t, and sj,t|j≠o are the share of consumers who purchase platform j, the outside good, and platform j conditional on purchasing a console at time t. Following prior literature, assuming a spot market for single-product software firms, free entry, and CES preferences for software, a reduced form equation relating the equilibrium number of software titles available to a console's installed base of users (IBj,t) can be derived: 28 (2) where ηj,t is a mean-zero error. Clements and Ohashi (2005) estimate these two equations across multiple generations of videogame consoles between 1994–2002 using price and quantity information provided by NPD Group, a market research firm (which also is the source for most of the market level data used in the majority of videogame papers discussed in this chapter). They employ console and year (p. 91) dummies in estimation, use the Japanese Yen and US dollar exchange rate and console retail prices in Japan as an instrument for price, and use the average age of software titles onboard each system as an instrument for the installed base. The main objects of interest are ω and γ in (1) and (2), which represent the responsiveness of consumer demand to the number of software titles, and vice versa. In Clements and Ohashi (2005) and similar studies, these coefficients are found to be significant and positive, which are interpreted as evidence of indirect network effects.29 Furthermore, these studies often show that such coefficients vary over time: e.g., Clements and Ohashi (2005) include age-interaction effects with installed base in (2), and find that the responsiveness of software to installed base decreases over lifetime of videogame console; similarly, Chintagunta, Nair, and Sukumar (2009) use an alternative hazard rate econometric specification of technology adoption and find that strength of network effects also varies over time, and that the number of software titles and prices have different effects on demand in later versus earlier periods. Both of these studies find price elasticities for a console diminish as consoles get older. It is worth stressing (as these papers have) that these estimates come from a static model, and care must be used when interpreting estimated parameters. There are several reasons a static model may not be ideal for analyzing this industry. Since consoles and games are durable goods, consumers do not repurchase the same product which typically is implied by a static model without inventory consideration; in addition, forward-looking consumers may delay purchase in anticipation of lower prices or higher utility from consumption in future periods (which may partially explain the strong seasonal spike in sales around the holidays). Failing to account for both the durability of goods and the timing of purchases can bias estimates of price and cross-price elasticities (c.f. Hendel and Nevo (2006)) as well as other parameters – including the strength of network effects. Most importantly, however, a static model does not allow consumers to anticipate future software releases when deciding when to purchase a console; since consoles are durable, consumers in reality base their hardware purchasing decisions on expectations over all software on a platform, including those titles that have not yet been released. Hence, a consumer's utility function for a console for example, should reflect this. Thus, insofar that ω can be estimated, it at best represents the extent to which current software variety reflects a consumer's expectation over the stock of current and future games. That estimated coefficients for these static models are shown to vary across time and even across consoles suggest dynamic issues are at play, and the underlying relationship between consumer demand and software availability may be significantly more complex. 3.1.1. Dynamics and Software/Consumer Heterogeneity In response to these concerns, researchers have begun incorporating dynamics into their analysis of consumer demand for videogames. For instance, Dubé, Hitsch, and Chintagunta (2010) utilize a dynamic model in which forward looking (p. 92) consumers time their purchases for consoles based on expectations of future prices and software availability; using a two-step estimator, they are also able to simultaneously estimate a console provider's optimal dynamic pricing function. Using estimates from their model, the authors study how indirect network effects Page 6 of 18 Home Videogame Platforms can lead to greater platform market concentration, and illustrate how, in spite of strong network effects, multiple incompatible platforms can co-exist in equilibrium. Nonetheless, this dynamic model of hardware demand still maintains the assumption used in the previous empirical network effects literature that consumers respond to software “variety,” which can be proxied by the number of available software titles, and that software variety can still be expressed as a reduced form function of each platform's installed base (e.g., as in (2)). This may have been a reasonable assumption for these papers which primarily focused on the period up until and including the 32/64 bit generation of videogames (roughly pre-2000). However, as mentioned previously, the past decade has seen the dominance of hit games where a small subset of software titles captured the majority of software sales onboard a console. Given the increasing variance in software quality and skewed distribution of software sales, a model specifying consumer utility as a function only of the number of software titles as opposed to the identity of individual games – although tractable and analytically convenient – may be of limited value in analyzing the most recent generations of the videogame industry as well as other “hit-driven” hardware-software markets. Mirroring the necessity to control for software heterogeneity in videogames is the additional need to control carefully for consumer heterogeneity. As has been previously discussed, the variance across consumers in the number of games purchased and hours spent playing games has been well documented, and capturing this rich heterogeneity is important for accurate estimates of product qualities and demand parameters. Although controlling for consumer heterogeneity is also important in a static setting, doing so in a dynamic context adds an additional complexity in that the characteristics of consumers comprising the installed base of a console evolves over time. E.g., since early adopters of videogame consoles are predominantly consumers with high valuations for videogames, software released early in a console's lifetime face a different population of consumers than a game that is released after a console is mature. Failing to correct for this consumer selection across time will bias upwards estimates of early-released games’ qualities, and bias downwards estimates of games released later. In turn, the magnitudes of these parameters underly incentives for almost all strategic decisions on the part of firms: e.g., firms may engage in intertemporal price discrimination (initially setting high prices before lowering them) in order to extract profits out of high valuation or impatient consumers first. In an attempt to control for these issues, Lee (2010a) estimates a dynamic structural model of consumer demand for both videogame hardware and software between 2000–2005 that explicitly incorporates heterogeneity in both consumer preferences and videogame quality. By explicitly modeling and linking hardware and software demand, the analysis is able to extract the marginal impact of a single (p. 93) videogame title on hardware sales, and allow this impact to differ across titles in an internally consistent manner. An overview of the approach follows. The model first specifies consumer i's lifetime expected utility of purchasing a hardware console j at time t (given she owns other consoles contained within her inventory l) as: (3) where xj,t are observable console characteristics, pj,t the console's price, ξj,t an unobservable product characteristic, and εi,j,t,ι a logit error. The paper introduces two additional terms to account for inventory concerns and the anticipation of future software: i.e., D(·) captures substitution effects across consoles and allows a consumer to value the console less (or more) if she already owns other consoles contained within ι; and Гj,t reflects a consumer's perception of the utility she would derive from being able to purchase videogames available today and in the future. Consumers have different preferences for videogaming, captured by the coefficient and for prices, given by , . Finally, the coefficient on Гj,t, α Г , captures how much hardware utility — and hence hardware demand – is influenced by expected software utility. Note that this specification of hardware demand does not use a static-period utility function, but rather lifetime expected utilities. Furthermore, the model incorporates dynamics explicitly by assuming consumers solve a dynamic programming problem when determining whether or not to purchase a videogame console in a given period: each consumer compares her expected value from purchasing a given console to the expected value from purchasing another console or none at all; furthermore, consumers can multihome and return the next period to purchase any console they do not already own. On the software side, the setup is similar: every consumer who owns a console is assumed to solve a similar Page 7 of 18 Home Videogame Platforms dynamic programming problem for each game she can play but does not already own. This in turn allows for the derivation of the expected option value of being able to purchase any particular title k onboard console j at time t, which is denoted EWi,j,k,t. Finally, to link the hardware and software demand together, the model defines Гj,t as the sum of option values for any software title k available on console j at time t (given by the set Kj,t) plus an expectation over the (discounted) option values of being able to purchase games to be released in the future, represented by Λj,t: (4) Lee (2010a) estimates the underlying structural parameters of the model, which include product fixed effects for every console and game and consumer (p. 94) preferences over price and software availability, utilizing techniques pioneered in Rust (1987), Berry (1994), and Berry, Levinsohn, and Pakes (1995), and later synthesized in a dynamic demand environment by Melnikov (2001) and Gowrisankaran and Rysman (2007).30 An important extension involves controlling for the selection of consumers onto consoles across time, which requires the simultaneous estimation of both hardware and software demand. Estimates indicate that although the vast majority of titles had a marginal impact on hardware demand, the availability of certain software titles could shift hardware installed bases by as much as 5 percent; furthermore, only a handful of such “hit” titles are shown to have been able to shift hardware demand by more than one percent over the lifetime of any given console. A model which assumed consumers valued all titles equally would thus lead to drastically different predictions on the impact and magnitudes of software on hardware demand. Lee (2010a) also demonstrates that by failing to account for dynamics, consumer heterogeneity, and the ability for consumers to purchase multiple hardware devices, predicted consumer elasticities with respect to price and software availability would be substantially biased. As is often the case, however, several strong assumptions are required for this more complicated analysis. First, for tractability, consumers perceive each software title onboard a system as an independent product.31 Second, consumers have rational expectations over a small set of state variables which are sufficient statistics for predicting future expected utilities. Although the consistency of beliefs with realized outcomes may have been a reasonable assumption for the period examined, there may be other instances for which it may not be well suited: e.g., Adams and Yin (2010) study the eBay resale of the newest generation of videogames consoles released in 2006, and find that prices for pre-sale consoles rapidly adjust after they are released.32 3.1.2. Software Supply and Pricing Accompanying the development of more realistic models for consumer demand have been richer models for software supply which treat software firms as dynamic and strategic competitors. One strain of literature focused on the optimal pricing of videogame software, itself a general durable goods market with forward looking consumers. Nair (2007) combines a model of dynamic consumer demand for videogame software with a model of dynamic pricing, and finds that the optimal pricing strategy for a software firm is consistent with a model of “skimming”: charging high prices early to extract rents from high value (or impatient) consumers before dropping prices over time to reach a broader market. This corresponds to the pricing patterns observed in the data: the vast majority of games on a console are released at a single price point (e.g., $49.99), and prices fall in subsequent periods.33 Inevitably, studies on pricing can only be conducted on games which have been already released for a particular platform; moving one step earlier in a software developer's decision process is the choice of which console to join. A first-party game has historically only been released exclusively on the integrated platform; (p. 95) however, a third-party software developer has a strategic choice: it can release a title on multiple platforms in order to reach a larger audience but pay additional porting costs, or it can develop exclusively for one console and forgo selling its game to consumers on other platforms. Lee (2010b) models what can be considered software's “demand” for a platform. As in consumer demand, Page 8 of 18 Home Videogame Platforms dynamics are important in this decision as well: since each software publisher makes this choice months before a game's release and since a game remains on the market for at least several months, a software developer anticipates changes in future installed bases of each console as well as the subsequent choices of other software developers when comparing expected profits of different actions. Using both the consumer demand estimates and similar assumptions used in Lee (2010a), the model computes a dynamic rational expectations equilibrium in which every software title chooses the optimal set of platforms to develop for while anticipating the future actions (and reactions) of other agents. A key input into the model, however, are porting costs for supporting different sets of consoles. These are typically unobserved. Lee (2010b) estimates these costs for games released between 2000 and 2005 under the revealed preference assumption that games released in the data were released on the subset of platforms which maximized their expected discounted profits.34 Via an inequalities estimator developed in Pakes, Porter, Ho, and Ishii (2006), relative differences in porting costs can be estimated. Estimates show significant variance in costs depending on the genre or type of game being ported, and that some consoles are cheaper (e.g., Xbox) than others (e.g., PS2) to develop for. On average, costs for this generation are approximately $1M per additional console, which are roughly in line with industry estimates. The final step back in the software production sequence involves the creation and development of new games. This represents the least developed area of research on software supply, and is the remaining key step in completely unpacking the mechanism which generates the reduced form relationship shown to exist between installed base of a console and software availability. On this front are issues related to an investment-quality tradeoff for game development, a product positioning decision of what genre or type of game to produce, timing games with release dates as with motion pictures (Einav (2007)), and the make-or-buy decision faced by a software publisher who can either engage in an arms-length contract with an independent developer or integrate into software development. Although some papers have studied whether integration with a console provider improves game quality,35 there remains much to be done. 3.2. Platform Competition Most of the analysis discussed so far has held fixed the actions of each platform, including choices of royalty rates charged to third-party software providers, development or porting costs, and integration or exclusive contracting decisions. (p. 96) Understanding these decisions from a theoretical perspective is complicated; the ability to analyze these strategic choices is further confounded by the absence of detailed data on these objects of interest (i.e., royalties, costs, and contracts). Even so, understanding how videogame platforms compete with one another for consumers and software firms is not only perhaps the most important aspect of this industry, but also the one that is the most relevant and generalizable to other hardware-software markets and platform industries. Thus overcoming these challenges should be the focus of future efforts. 3.2.1. Pricing The most developed area of research on platform strategy has been on the pricing of hardware consoles: both Dubé, Hitsch, and Chintagunta (2010) and Liu (2010) estimate dynamic models of hardware pricing to consumers, and highlight the importance of indirect network effects in explaining observed pricing patterns and rationalizing console “penetration pricing” – that is, consoles are typically priced below marginal costs, but as marginal costs fall faster than prices, margins tend to increase over time. Dubé, Hitsch, and Chintagunta (2010) further note that the presence of network effects are not sufficient on their own to make penetration pricing optimal, and rather that these effects need be sufficiently strong. Nonetheless, these analyses hold fixed prices charged by platforms to the “other side” of the market in that the supply of software is only dependent on the installed base of consumers onboard a console, and not the royalty rates levied by the console. Of course, in reality these royalties are as much a strategic decision as the price charged to consumers, and in many ways are just as important to a platform's success. For example, Sony charged a much lower royalty than Nintendo when it introduced its Playstation console ($9 as opposed to $18), which helped it attract a greater number of third-party software developers (Coughlan, 2001). To determine the optimal royalty, it's useful to understand why they need be positive at all. The theoretical two- Page 9 of 18 Home Videogame Platforms sided market literature (c.f. Armstrong (2006); Rochet and Tirole (2006); Weyl (2010)) has focused on precisely this question in related networked industries, and emphasized how changing the division of pricing between sides of a platform market can affect platform demand and utilization; Hagiu (2006) focused on the videogame industry in particular, and analyzed the relationship between a console's optimal royalty rate and optimal hardware price. As noted before, the videogame industry differs from most other hardware-software markets such as the PC industry in that the majority of platform profits derive not from the end user or consumer, but rather from the software developers in the form of royalty payments. However, providing a single explanation of why this occurs within the videogame industry proves difficult, as many theory models indicate which side can multihome, how much one side responds and values the participation of the other, and the heterogeneity in such preferences can drastically influence the optimal division of prices.36 Thus, there may be many forces at work; Hagiu (2009) provides another explanation, in which the more that consumers (p. 97) prefer a variety of software products, the greater a platform's profits derive from software in equilibrium. There are difficulties testing these alternative explanations in the data. First, although obtaining measurements of elasticities of consumers with respect to software (and vice versa) is possible, estimating how software supply would change in response to a change in royalty rates is difficult; not only is data on royalty rates difficult to come by, but typically they do not change for a particular console during its lifetime and hence there is little identifying variation.37 Second, given that certain hit software titles dominate the market and games are supplied increasingly by publishers with market power, it is an open question whether the theoretical results still apply when one side of the market no longer are price-takers but rather strategic oligopolists. 3.2.2. Porting Costs and Compatibility Another decision on the part of console manufacturers that has not widely been studied is the ability of a platform provider to affect the costs of developing or “porting” to its console. The theoretical literature has studied the role of switching costs in influencing market share and power in general networked industries (c.f. Farrell and Klemperer (2007)), and these issues are central in the videogame market as well. For example, anecdotal evidence suggests that one of the main reasons for Sony's success in entering the videogame market was that it was easier and cheaper to develop for the Sony Playstation than rival consoles at the time: in addition to having lower royalty rates, Sony actively provided development tools and software libraries to third party developers, and it utilized CDs as opposed to more costly cartridges (the format used by Nintendo consoles at the time). Incidentally, Microsoft also leveraged lower development costs as a selling point of its first console: as essentially a modified Microsoft Windows PC with an Intel CPU, the Xbox was extremely easy for existing PC developers to adjust to and develop games for (Takahasi, 2002). Relatedly, platform providers can also decide whether or not to make games compatible across consoles, as opposed to forcing developers to make different versions. Although cross-platform compatibility across competing consoles has not been witnessed (instead, requiring software developers to port and create multiple versions of a game), a platform provider could allow for backwards compatibility – that is, a new console being able to play games released for the previous generation console. One widely cited advantage of Sony's PS2 over its competitors at the time was its compatibility with original Playstation games; this gave it an accessible library of over a thousand games upon release, easily surpassing the number of playable titles on any of its competitors.38 Interestingly, the PS3 initially could play PS2 games, but newer versions of the console eliminated this ability; this suggests that the benefits to backward compatibility are most substantial early in a console's life before currentgeneration games are widely available, and later may not be worth the cost.39 (p. 98) 3.2.3. Exclusivity and Integration Although there is some degree of hardware differentiation across consoles, the primary means by which consoles compete with one another for consumers (in addition to price) is through the availability of exclusive games.40 Before Sony entered the video-game business in 1993 with its Playstation console, it purchased a major software developer in addition to securing agreements for several exclusive titles (Kent, 2001). Similarly, before launching the Xbox in 2001, Microsoft bought several software developers to produce exclusive games; many attribute the (relative) success of Microsoft's Xbox console to its exclusive game Halo, acquired in 2000. In both instances, having high-quality games available with the release of a console that were not available on competitors contributed to greater sales. Page 10 of 18 Home Videogame Platforms A platform typically obtains an exclusive game in one of two ways: via internal development by a integrated developer, or via payment to a third party developer. In recent years as development costs for games have been increasing and porting costs have fallen as a percentage of total costs, most third-party titles have chosen to multihome and support multiple consoles in order to maximize their number of potential buyers. Thus, even though exclusive arrangements still occur for third-party titles, they are now increasingly used for only a temporary period of time (e.g., six months), and console providers have become even more reliant on their own first-party titles to differentiate themselves. In general, understanding how platforms obtain exclusive content – either via integration or exclusive contracting – requires a model of bilateral contracting with externalities between console manufacturers and software developers (c.f. Segal (1999); Segal and Whinston (2003); de Fontenay and Gans (2007)). For example, the price Sony would need to pay a software developer for exclusivity depends crucially on how much Sony would benefit, as well as how much Sony would lose if Microsoft obtained exclusivity over the title instead. Unfortunately, the applicability of theory to settings with multiple agents on both sides of the market is limited (there are at least three major console manufacturers and multiple software publishers and developers), and is even further confounded by the presence of dynamics.41 Although static models of bargaining for exclusivity have been analyzed,42 a general model that can be taken to the data and inform our ability to predict which games or developers would be exclusive, and the determinants of the negotiated price, would be extremely useful.43 3.2.4. Other Concerns Ultimately, one of the biggest hurdles in bringing the theory to the data may very well be identifying the incentives each major platform provider faces. Both Sony and Microsoft have multiple other platform businesses which are affected by decisions made within their videogame divisions. For example, Sony faced a much higher marginal cost than its competitors as a result of including its proprietary Blu-ray player in its PS3 console; such a decision was a response to its desire to win the standards battle for a next-gen DVD format over consumer electronics rival Toshiba. (p. 99) In addition, Microsoft viewed the Xbox as partly a means of protecting and expanding its Windows and PC software business during an era of digital convergence (Takahasi, 2002). In both cases, each company sustained large initial losses in their videogame divisions ($4B in the first four years of the Xbox for Microsoft, $3.3B for Sony in the first two years of its PS3),44 but focusing on these numbers alone would understate the total benefits each company received.45 Furthermore, there is again a dynamic aspect: had Microsoft not entered in 2000 with a viable platform, it would have had a more difficult time releasing its Xbox360 device in 2005. Determining the appropriate scope across industries and time horizon each company faces when making strategic decisions is an open challenge. 3.3. Vertical Issues 3.3.1. Exclusive Software for Consoles The forced exclusivity contracts employed by Nintendo in the 1980's – whereby developers could only develop exclusively for Nintendo or not at all – were dropped under legal and regulatory pressure in 1990. Since then, many have argued that these were anticompetitive not only in videogames (e.g., Shapiro (1999)), but in other industries (e.g., U.S. v. Visa) as well. Nonetheless, exclusive games persist. A natural question, thus, is whether the continued presence of exclusive first-party games developed internally by platforms, or the use of lump-sum payments by platforms in exchange for exclusivity from third-party software developers, can be anticompetitive. Theory has shown the effects of such exclusive vertical relationships can be ambiguous. Such relationships can be used to deter entry or foreclose rivals ((Mathewson and Winter (1987), Rasmusen, Ramseyer, and Wiley (1991), Bernheim and Whinston (1998)), which may be exacerbated by the presence of network externalities (e.g., Armstrong and Wright (2007)).46 Furthermore, exclusivity can limit consumer choice and hence welfare by preventing consumers on competing platforms from accessing content, products, or services available only elsewhere. On the other hand, exclusive arrangements may have pro-competitive benefits, such as encouraging investment and effort provision by contracting partners ((Marvel (1982), Klein (1988), Besanko and Perry (1993), Segal and Whinston (2000)). In networked industries, integration by a platform provider may be effective in solving the “chicken-and-egg” coordination problem, one of the fundamental barriers to entry discussed in the two-sided market literature. Furthermore, exclusivity may be an integral tool used by entrant platforms to break into Page 11 of 18 Home Videogame Platforms established markets: by preventing contracting partners from supporting the incumbent, an entrant can gain a competitive advantage, spur adoption of its own platform, and thereby spark greater platform competition. Both Prieger and Hu (2010) and Lee (2010b) attempt to shed light on this question in the context of the sixth generation of the videogame industry (2000–2005). (p. 100) Prieger and Hu (2010) use a demand model similar to Clements and Ohashi (2005) to show that the marginal exclusive game does not affect console demand; consequently, the paper suggests that a dominant platform cannot rely on exclusive titles to dominate the market. However, as already discussed in this chapter, controlling for heterogeneity in game quality is crucial, and cannot be captured in a model where consumers only value the number of software products: estimates from Lee (2010a) show that games that actually could drastically affect hardware market shares were primarily integrated or exclusively provided to only one console. Thus, insofar the few hit games onboard the largest platform of the time period studied could have contributed to its dominant position, exclusive vertical arrangements may have led to increased market concentration. To explore this possibility, Lee (2010b) conducts a counterfactual environment in which exclusive vertical arrangements were prohibited in the industry during the time period studied: that is, all hardware providers both could not write exclusive software or write exclusive contracts with software providers. Using the techniques described in the previous chapter and demand estimates from Lee (2010a), Lee (2010b) simulates forward the market structure if all consumers and games (including those that previously had been integrated) could freely choose which platforms to purchase or join, and solves for the dynamic equilibrium of this game. The main finding, focusing on the platform adoption decisions of consumers and software, is that banning exclusive arrangements between hardware platforms and software publishers would have actually benefited Sony, the dominant “incumbent” platform (with the one-year head start and larger installed base), and harmed the two smaller platforms (Microsoft and Nintendo) during the time period studied. The intuition for this result is straightforward: without exclusive arrangements, the developers of high quality software would typically multihome and support all three consoles; lower quality titles, constrained by the costs of porting, would likely develop first for the incumbent due to its larger installed base, and only later, if at all, developed a version for either entrant platform. As a result, neither entrant platform would have been able to offer consumers any differentiation or benefit over the incumbent. With exclusivity, however, entrants could create a competitive advantage, and was hence leveraged by them to gain traction in this networked industry.47 The paper still notes that even though banning exclusive vertical arrangements may have increased industry concentration, consumers may have benefited from access to a greater selection of software titles onboard any given platform: consumer welfare would have increased during the five-year period without exclusivity since a consumer could access a greater number of software titles while needing to only purchase one console. Nonetheless, the analysis abstracts away from many potential responses to such a counterfactual policy change: for example, platform providers are assumed to offer the same non-discriminatory contracts to all firms, investment and product qualities do not change, and prices, entry, and exit of all products are held fixed. Indeed, the paper notes that if Sony's prices increased as (p. 101) a result of its increased market share (or if either Nintendo or Microsoft exited in that generation or the subsequent one, or software supply was adversely affected by the efficiency benefits of integration and exclusivity) the change to consumer welfare could easily have been significantly negative. Thus, although it does appear that both Microsoft and Nintendo benefited from the ability to engage in exclusive dealing in this period, the effects on consumer welfare are ambiguous; furthermore, in order to paint a complete story of the effects of integration or exclusivity, one might also wish to examine an environment in which only certain platforms (e.g., the incumbent or dominant player) could not engage in exclusive contracting, but others could. Such extensions would require developing additional tools to analyze the broader set of strategic decisions facing software and hardware firms discussed previously. 4. Concluding Remarks The home videogame market is but a portion of the entire videogame industry, yet has proven to be a rich testing ground for models of strategic interaction and theories of platform competition. The literature that has developed, though still nascent, has shown the potential for tackling and addressing myriad issues simply by studying an Page 12 of 18 Home Videogame Platforms industry which once was considered just a curiosity and fad. Looking forward, the continued growth of the videogame industry has the potential for being both a curse and a boon for research. On one hand, as videogames become even more pervasive and intertwined with other industries, it becomes – to a certain degree – less suited for “clean” and tractable analysis. Indeed, one of the advantages of studying the home videogame market was precisely the relative ease in which the relevant agents and parties could be identified; going forward, this no longer may be the case. Furthermore, as this chapter discussed, even retrospective analysis of the videogame market when all the players were known required sophisticated modeling techniques to handle the industry's complexities, which include controlling for dynamic issues and accounting for consumer and software heterogeneity. Accounting for even more complicated strategic interactions poses a daunting challenge. On the other hand, the success and proliferation of videogames will continue to spawn broader questions and improve our understanding of general networked industries. At the heart of digital convergence is the story of dominant platforms in once separated markets suddenly finding themselves to be competitors: much as videogame consoles encroach upon adjacent markets such as content distribution, so to have firms in other markets – for example, smartphone manufacturers, social networking platforms – ventured into the gaming business. How this cross-industry platform competition will play out and adapt to changing environments remains a fascinating topic for exploration. References Adams, C., and P.-L. Yin (2010): “Reallocation of Video Game Consoles on eBay,” mimeo. Armstrong, M. (2006): “Competition in Two-Sided Markets,” RAND Journal of Economics, 37(3), 668–691. Armstrong, M., and J. Wright (2007): “Two-Sided Markets, Competitive Bottlenecks, and Exclusive Contracts,” Economic Theory, 32(2), 353–380. Bernheim, B. D., and M. D. Whinston (1998): “Exclusive Dealing,” Journal of Political Economy, 106(1), 64–103. Berry, S. (1994): “Estimating Discrete-Choice Models of Product Differentiation,” RAND Journal of Economics, 25(2), 242–262. Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile Prices in Market Equilibrium,” Econometrica, 63(4), 841– 890. Besanko, D., and M. K. Perry (1993): “Equilibrium Incentives for Exclusive Dealing in a Differentiated Products Oligopoly,” RAND Journal of Economics, 24, 646–667. Brandenburger, A. (1995): “Power Play (C): 3DO in 32-bit Video Games,” Harvard Business School Case 795–104. (p. 105) Chintagunta, P., H. Nair, and R. Sukumar (2009): “Measuring Marketing-Mix Effects in the Video-Game Console Market,” Journal of Applied Econometrics, 24(3), 421–445. Chou, C., and O. Shy (1990): “Network Effects without Network Externalities,” International Journal of Industrial Organization, 8, 259–270. Church, J., and N. Gandal (1992): “Network Effects, Software Provision, and Standardization,” Journal of Industrial Economics, 40(1), 85–103. Clements, M. T., and H. Ohashi (2005): “Indirect Network Effects and the Product Cycle: U.S. Video Games, 1994 – 2002,” Journal of Industrial Economics, 53(4), 515–542. Corts, K. S., and M. Lederman (2009): “Software Exclusivity and the Scope of Indirect Network Effects in the US Home Video Game Market,” International Journal of Industrial Organization, 27(2), 121–136. Page 13 of 18 Home Videogame Platforms Coughlan, P. J. (2001): “Note on Home Video Game Technology and Industry Structure,” Harvard Business School Case 9-700-107. Dranove, D., and N. Gandal (2003): “The DVD vs. DIVX Standard War: Empirical Evidence of Network Effects and Preannouncement Effects,” Journal of Economics and Management Strategy, 12, 363–386. Dubé, J.-P., G. J. Hitsch, and P. Chintagunta (2010): “Tipping and Concentration in Markets with Indirect Network Effects,” Marketing Science, 29(2), 216–249. Einav, L. (2007): “Seasonality in the U.S. Motion Picture Industry,” RAND Journal of Economics, 38(1), 127–145. Eisenmann, T., and J. Wong (2005): “Electronic Arts in Online Gaming,” Harvard Business School Case 9-804-140. Evans, D. S., A. Hagiu, and R. Schmalensee (2006): Invisible Engines: How Software Platforms Drive Innovation and Transform Industries. MIT Press, Cambridge, MA. Farrell, J., and P. Klemperer (2007): “Coordination and Lock-In: Competition with Switching Costs and Network Effects,” in Handbook of Industrial Organization, ed. by M. Armstrong, and R. Porter, vol. 3. North-Holland Press, Amsterdam. Farrell, J., and G. Saloner (1986): “Installed Base and Compatibility: Innovation, Product Preannouncements, and Predation,” American Economic Review, 76, 940–955. Gandal, N., M. Kende, and R. Rob (2000): “The Dynamics of Technological Adoption in Hardware/Software Systems: the Case of Compact Disc Players,” RAND Journal of Economics, 31, 43–61. Gil, R., and F. Warzynski (2009): “Vertical Integration, Exclusivity and Game Sales Performance in the U.S. Video Game Industry,” mimeo. Gowrisankaran, G., and M. Rysman (2007): “Dynamics of Consumer Demand for New Durable Goods,” mimeo. Hagiu, A. (2006): “Pricing and Commitment by Two-Sided Platforms,” RAND Journal of Economics, 37(3), 720–737. —— (2009): “Two-Sided Platforms: Product Variety and Pricing Structures,” Journal of Economics and Management Strategy, 18(4), 1011–1043. Hagiu, A., and R. S. Lee (2011):”Exclusivity and Control,” Journal of Economics and Management Strategy, 20(3), 679–708. Hendel, I., and A. Nevo (2006): “Measuring the Implications of Sales and Consumer Inventory Behavior,” Econometrica, 74(6), 1637–1673. Katz, M., and C. Shapiro (1985): “Network Externalities, Competition, and Compatibility,” American Economic Review, 75, 424–440. (p. 106) Kent, S. L. (2001): The Ultimate History of Video Games. Three Rivers Press, New York, NY. Klein, B. (1988): “Vertical Integration as Organizational Ownership: The Fisher Body-General Motors Relationship Revisited,” Journal of Law, Economics and Organization, 4, 199–213. Lee, R. S. (2010a): “Dynamic Demand Estimation in Platform and Two-Sided Markets,” mimeo, http://pages.stern.nyu.edu/~rslee/papers/DynDemand.pdf. —— (2010b): “Vertical Integration and Exclusivity in Platform and Two-Sided Markets,” mimeo, http://pages.stern.nyu.edu/~rslee/papers/VIExclusivity.pdf. Lee, R. S., and K. Fong (2012): “Markov-Perfect Network Formation: An Applied Framework for Bilateral Oligopoly and Bargaining in Buyer-Seller Networks” mimeo, http://pages.stern.nyu.edu/~rslee/papers/MPNENetworkFormation.pdf Liu, H. (2010): “Dynamics of Pricing in the Video Game Console Market: Skimming or Penetration?,” Journal of Page 14 of 18 Home Videogame Platforms Marketing Research, 47(3), 428–443. Marvel, H. P. (1982): “Exclusive Dealing,” Journal of Law and Economics, 25, 1–25. Mathewson, G. F., and R. A. Winter (1987): “The Competitive Effects of Vertical Agreements,” American Economic Review, 77, 1057–1062. Melnikov, O. (2001): “Demand for Differentiated Products: The case of the U.S. Computer Printer Market,” mimeo. Nair, H. (2007): “Intertemporal Price Discrimination with Forward-looking Consumers: Application to the US Market for Console Video-Games,” Quantitative Marketing and Economics, 5(3), 239–292, forthcoming. Nair, H., P. Chintagunta, and J.-P. DubE (2004): “Empirical Analysis of Indirect Network Effects in the Market for Personal Digital Assistants,” Quantitative Marketing and Economics, 2(1), 23–58. Ohashi, H. (2003): “The Role of Network Effects in the U.S. VCR Market: 1978–1986,” Journal of Economics and Management Strategy, 12, 447–496. Pakes, A., J. Porter, K. Ho, and J. Ishii (2006): “Moment Inequalities and Their Application,” mimeo, Harvard University. Prieger, J. E., and W.-M. Hu (2010): “Applications Barriers to Entry and Exclusive Vertical Contracts in Platform Markets,” Economic Inquiry, forthcoming. Rasmusen, E. B., J. M. Ramseyer, and J. S. Wiley (1991): “Naked Exclusion,” American Economic Review, 81(5), 1137–1145. Rey, P., and J. Tirole (2007): “A Primer on Foreclosure,” in Handbook of Industrial Organization, ed. by M. Armstrong, and R. Porter, vol. 3. North-Holland Press, Amsterdam. Riordan, M. H. (2008): “Competitive Effects of Vertical Integration,” in Handbook of Antitrust Economics, ed. by P. Buccirossi. MIT Press, Cambridge, MA. Rochet, J.-C., and J. Tirole (2006): “Two-Sided Markets: A Progress Report,” RAND Journal of Economics, 37(3), 645–667. Rust, J. (1987): “Optimal replacement of GMC bus engines: An empirical model of Harold Zurcher,” Econometrica, 55, 999–1033. Segal, I. (1999): “Contracting with Externalities,” Quarterly Journal of Economics, 64(2), 337–388. Segal, I., and M. D. Whinston (2000): “Exclusive Contracts and Protection of Investments,” RAND Journal of Economics, 31, 603–633. —— (2003): “Robust Predictions for Bilateral Contracting with Externalities,” Econo-metrica, 71(3), 757–791. (p. 107) Shankar, V., and B. L. Bayus (2003): “Network Effects and Competition: An Empirical Analysis of the Home Video Game Market,” Strategic Management Journal, 24(4), 375384. Shapiro, C. (1999): “Exclusivity in Network Industries,” George Mason Law Review, 7, 673–683. Stennek, J. (2007): “Exclusive Quality – Why Exclusive Distribution May Benefit the TV Viewers,” IFN WP 691. Takahasi, D. (2002): Opening the Xbox. Prima Publishing, Roseville, CA. Weyl, E. G. (2010): “A Price Theory of Multi-Sided Platforms,” American Economic Review, 10(4). Whinston, M. D. (2006): Lectures on Antitrust Economics. MIT Press, Cambridge, MA Notes: Page 15 of 18 Home Videogame Platforms (1.) DFC Intelligence, 2009. 2010 Essential Facts, Entertainment Software Association. (2.) Ibid. (3.) Ibid. (4.) 2010 Media Industry Fact Sheet, Nielsen Media Research. (5.) Sony also released a general purpose computer system called the MSX in 1983 that could be used for gaming. (6.) http://www.scei.co.jp/corporate/release/100118e.html. (7.) 2010 Media Industry Fact Sheet, Nielsen Media Research. (8.) “Natal vs. Sony Motion Controller: is the console cycle over?” guardian.co.uk, February 26, 2010. (9.) ESA 2010 Essential Facts. (10.) Originally, the first videogame consoles were essentially integrated hardware and software devices provided by a single firm; not until after 1976, with the release of the Fairchild System F and the Atari VCS, were other firms able to produce software for videogame consoles via the use of interchangeable cartridges. (11.) Average costs reached $6M during the late 1990s (Coughlan, 2001), and today can range between $20 – 30M for the PS3 and Xbox360 (“The Next Generation of Gaming Consoles,” CNBC.com, June 12, 2009). (12.) E.g., “Call of Duty: Modern Warfare 2 tops $1 billion in sales,” Los Angeles Times, January 14, 2010. (13.) 3DO was a console manufacturer who tried a different pricing scheme by significantly marking up its hardware console, but charging no software royalties. 3DO survived for 3 years before filing for bankruptcy in 2003 (Brandenburger, 1995). (14.) See Hagiu, “Software Platforms,” chapter 3 in this Handbook for more discussion. (15.) “Will Xbox drain Microsoft?,” CNET News, March 6, 2001. “Microsoft's Midlife Crisis,” Forbes, September 13, 2005. (16.) “Sony Eyes Return to Profit,” Wall Street Journal, May 14, 2010. (17.) “Nintendo takes on PlayStation, Xbox,” Reuters, September 14, 2006. (18.) A notable exception is “backwards compatibility,” which refers to the ability of a new console to use software developed for the previous version of that particular console. E.g., the first version of the PS3 could play PS2 games, and the PS2 could play PS1 games; the Xbox360 can play Xbox games. (19.) Industry sources; Eisenmann and Wong (2005) cite $1M as the porting cost for an additional console for the sixth generation of platforms. (20.) “Nintendo to ease restrictions on U.S. game designers,” The Wall Street Journal, October 22, 1990. Kent (2001). (21.) This may also partially be a result of the fact that no console has since matched Nintendo's 80–90 percent market share achieved in the late 1980s. (22.) ESA Essential Facts, 2010. (23.) “Video Game Culture: Leisure and Play Preferences of B.C. Teens,” Media Awareness Network, 2005. (24.) “Digital Gaming in America Survey’ Results,” Gaming Age, August 12, 2002. (27.) Whether the negative competition effect between two substitutable games dominates this positive network effect depends on the relative elasticities for adoption, which in turn typically depends on how early it is in a console's lifecycle. Page 16 of 18 Home Videogame Platforms (28.) See also Dubé, Hitsch, and Chintagunta (2010) for the derivation of a similar estimating equation. (29.) Corts and Lederman (2009) also find evidence of “cross-platform” network effects from 1995 to 2005: i.e., given the ability of software to multihome, software supply for one console was shown to be responsive to the installed bases across all platforms; as a result, users on one console could benefit from users on another incompatible console in that their presence would increase software supply for all consoles. (30.) Estimation of the model follows by matching predicted market shares for each hardware and software product over time from the model with those observed in the data (obtained from the NPD Group), and minimizing a GMM criterion based on a set of conditional moments. The main identifying assumption is that every product's one dimensional unobservable characteristic (for hardware, represented by ξ in (3)) evolves according to an AR(1) process, and innovations in these unobservables are uncorrelated with a vector of instruments. (31.) Since videogames are durable goods, keeping track of each consumers’ inventory and subsequent choice sets for over 1500 games was computationally infeasible. However, both Nair (2007) and Lee (2010a) provide evidence which suggests independence may not be unreasonable for videogames. (32.) Whether or not consumer beliefs can be estimated or elicited without imposing an assumption such as rational expectations is an important area of research for dynamic demand estimation in general. (33.) Nair (2007) provides anecdotal evidence that managers follow rules-of-thumb pricing strategies in which prices are revised downward if sales are low for a game, and keep prices high if sales are high. There is also evidence that consumers prefer newer games over older ones (e.g., Nair (2007) and Lee (2010a) both find significant decay effects in the quality of a game over time). (34.) The analysis ignores games that are contractually exclusive, which are discussed later in this chapter; it furthermore assumes publishers maximize profits individually for each game. (35.) E.g., Gil and Warzynski (2009) study videogames released between 2000 and 2007 and find reduced form evidence that indicates once release timing and marketing strategies are controlled for, vertically integrated games are not of higher quality than non-integrated games. However, regressions on the software fixed effects recovered in Lee (2010a) for a similar time period show first-party games are generally of higher quality. (36.) See also Evans, Hagiu, and Schmalensee (2006) for discussion. (37.) Appealing to cross-platform variation in royalty rates would require considerable faith that other console specific factors affecting software supply can be adequately controlled for. (38.) Nintendo and Microsoft followed suit with their seventh generation consoles. (39.) The original PS3 console included the PS2 graphic chip, which was eliminated in subsequent versions. (40.) Clearly, any game that multihomes and is available on multiple systems yields no comparative advantage across consoles. (41.) For instance, the gains to exclusivity depend on the age of the console (among other things), and platforms may choose to divest integrated developers later. E.g., Microsoft acquired the developer Bungie prior to launch of original Xbox in 2000; in 2007, it was spun off as Microsoft reasoned Bungie would be more profitable if it could publish for other consoles (“Microsoft, ‘Halo’ maker Bungie split,” The Seattle Times, October 6, 2007). (42.) For example, Hagiu and Lee (2011) apply the framework of Bernheim and Whinston (1998) to analyze exclusive contracting in platform industries; see also Stennek (2007). (43.) See Lee and Fong (2012) for progress along these lines. (44.) “Microsoft's Midlife Crisis,” Forbes, September 13, 2005; “PlayStation Poorhouse,” Forbes, June 23, 2008. (45.) Further confounding matters are each console manufacturer's online gaming businesses; Microsoft's online service generates over $1B a year (“Microsoft's Online Xbox Sales Probably Topped $1 Billion,” Bloomberg, July 7, 2010), and all 3 current-generation platforms have downloadable gaming stores as well. Page 17 of 18 Home Videogame Platforms (46.) Whinston (2006), Rey and Tirole (2007), and Riordan (2008) overview the theoretical literature on vertical foreclosure and the competitive effects of exclusive vertical arrangements. (47.) Note that had Sony's exclusive titles been significantly higher quality than those onboard Microsoft's or Nintendo's consoles, this result may have been different: i.e., even though the two entrant platforms would have lost their exclusive titles, they would have gained access (albeit non-exclusively) to Sony's hit exclusive titles. Nevertheless, demand estimates clearly indicate this was not the case. The question of how Nintendo and Microsoft were able to get access to higher quality software in the first place is beyond the scope of the paper, as it requires addressing questions raised in the previous section regarding software supply and hardware-software negotiations. Robin S. Lee Robin S. Lee is Assistant Professor of Economics at the Stern School of Business at New York University. Page 18 of 18 Digitization of Retail Payments Oxford Handbooks Online Digitization of Retail Payments Wilko Bolt and Sujit Chakravorti The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0005 Abstract and Keywords This article reports the research on electronic payment systems. Debit, credit, and prepaid are three forms of payment card. The rapid growth in the use of electronic payment instruments, especially payment cards, is a striking feature of most modern economies. Payment data indicate that strong scale economies exist for electronic payments. Payment costs can be decreased through consolidation of payment processing operations to realize economies of scale. Competition does not necessarily enhance the balance of prices for two-sided markets. The ability of merchants to charge different prices is a powerful incentive to convince consumers to use a certain payment instrument. The effect of interventions in Australia, Spain, the European Union, and the United States is dealt with. The theoretical literature on payment cards continues to grow. However, there are a few areas of payment economics that deserve greater attention. Keywords: debit cards, credit cards, prepaid cards, payment economics, payment cards, Australia, Spain, European Union, United States 1. Introduction Rapid advancements in computing and telecommunications have enabled us to interact with each other digitally. Instead of visiting a travel agent for information regarding our next vacation, we can purchase our vacation package online in the middle of the night. Prior to boarding, we can purchase and download a book to our Kindle to read during the flight. We no longer have to return home to share our vacation experience with our friends and family but can instead share our digital pictures taken with our iPhone via email or post them on Facebook. Despite the digital economy being upon us, we still rely on paper payment instruments such as cash, checks, and paper giros for a significant amount of face-to-face and remote bill payments in advanced economies. While we have not attained the cashless society, we have made significant strides to adopt electronic payment instruments. The proliferation of payment cards continues to change the way consumers shop and merchants sell goods and services. Recently, some merchants have started to accept only card payments for safety and convenience reasons. For example, several US airlines only accept payment cards for inflight purchases on all their domestic routes. Also, many quick service restaurants and coffee shops now accept payment cards to capture greater sales and increase transaction speed. Wider acceptance and usage of payment cards suggest that a growing number of (p. 109) consumers and merchants prefer payment cards to cash and checks. Furthermore, without payment cards, Internet sales growth would have been substantially slower. The increased usage of cards has increased the value of payment networks, such as Visa Inc., MasterCard Worldwide, Discover Financial Services, and others. In 2008, Visa Inc. had the largest initial public offering (IPO) of equity, valued at close to $18 billion, in US history (Benner, 2008). The sheer magnitude of the IPO suggests that financial market participants value Visa's current and future profitability as a payment network. Page 1 of 21 Digitization of Retail Payments Over the last decade or so, public authorities have questioned the underlying fee structures of payment networks and often intervened in these markets.1 The motivation to intervene varies by jurisdiction. Public authorities may intervene to improve the incentives to use more efficient payment instruments. For example, they may encourage electronic payments over cash and checks. Public authorities may also intervene because fees are “too high.” Finally, public authorities may enable adoption of payment standards that may be necessary for market participants to invest in new payment instruments and channels especially during times of rapid innovation and competing standards. In this chapter, we emphasize regulation of certain prices in the retail payment system. To date, there is still little consensus—either among policymakers or economic theorists—on what constitutes an efficient fee structure for payments. There are several conclusions that we draw from the academic literature. First, there are significant scale economies and likely scope economies in attracting consumers and merchants and payment processing. Second, cross-subsidies between consumers and merchants may be socially optimal suggesting that there are benefits to having a limited number of networks. Third, allowing merchants to price differentiate among different types of payment instruments generally more accurately reflect underlying costs to all participants. Fourth, merchant, card issuer, or network competition may result in lower social welfare contrary to standard economic principles. Finally, public authorities should not only consider the costs of payment processing but also consider the benefits received by consumers and merchants, such as convenience, security, and access to credit that may result in greater sales if they choose to intervene in payments markets. The rest of our article is organized as follows. We first explain the structure of payment networks. Having established a framework, we discuss consumer choice and the migration to electronic payments. Next, we describe the provision of payment services emphasizing the economies of scale and scope that are generally present. Then, we summarize the key contributions to the theoretical payment card literature focusing on economic surplus and cross subsidies and their impact on social welfare. In the following section, we discuss several market interventions by public authorities. Finally, we offer some concluding remarks and suggest future areas for research. (p. 110) 2. Structure of Payment Markets When focusing on payment cards, most card transactions occur in three- or four-party networks.2 These networks comprise of consumers and their banks (known as issuers), as well as merchants and their banks (known as acquirers). Issuers and acquirers are a part of a network that sets the rules and procedures for clearing and settling payment card receipts among its members. In principle, other forms of electronic payments, such as credit transfers and direct debits, have the same three or four party structure. Figure 5.1 Payment Card Fees. Source: Bolt and Chakravorti (2008b). In Figure 5.1, we diagram the four participants and their interactions with one another. First, a consumer establishes a relationship with an issuer and receives a payment card. Consumers often pay annual membership fees to their issuers. They generally do not pay per transaction payment card fees to their banks. On the contrary, some Page 2 of 21 Digitization of Retail Payments payment card issuers, usually more common for credit card issuers, give their customers per transaction rewards, such as cash back or other frequentuse rewards. Second, a consumer makes a purchase from a merchant. Generally, the merchant charges the same price regardless of the type of payment instrument used to make the purchase. Often the merchant is restricted from charging more for purchases that are made with payment cards. These rules are called no-surcharge rules.3 Third, if a merchant has established a relationship with an acquirer, it is able to accept payment card transactions. The merchant either pays a fixed per transaction fee (more common for debit cards) or a proportion of the total purchase amount, known as the merchant discount fee (more common for credit cards), to its acquirer.4 For credit cards, the merchant discount can range from one percent to five percent depending on the type of transaction, type of merchant, and (p. 111) type of card, if the merchant can swipe the physical card or not, and other factors. Fourth, the acquirer pays an interchange fee to the issuer. Debit, credit, and prepaid cards are three forms of payment cards. Debit cards allow consumers to access funds at their banks to pay merchants; these are sometimes referred to as “pay now” cards because funds are generally debited from the cardholder's account within a day or two of a purchase.5 Credit cards allow consumers to access lines of credit at their banks when making payments and can be thought of as “pay later” cards because consumers pay the balance at a future date.6 Prepaid cards can be referred to as “pay before” cards because they allow users to pay merchants with funds transferred in advance to a prepaid account.7 3. Payment Choice and Migration to Electronic Payments The rapid growth in the use of electronic payment instruments, especially payment cards, is a striking feature of most modern economies. In Table 5.1, we have listed the annual per capita payment transactions for ten advanced economies in 1988 and 2008. In all cases, there was tremendous growth, but countries differ significantly from one another and over time. For example, Italy had 0.33 per capita annual payment card transactions and the United States had 36.67 per capita annual payment card transactions in 1988 and 23.5 per capita annual payment card transactions and 191.1 per capita annual payment card transactions, respectively, in 2008. Also note that differences within Europe remain large. Countries like Italy, Germany, and Switzerland still have a strong dependence on cash use, whereas countries like the Netherlands, France, and Sweden show high annual per capita payment card volumes. Amromin and Chakravorti (2009) find that greater usage of debit cards resulted in lower demand for small-denomination banknotes and coins that are used to make change although demand for large-denomination notes has not been affected. From Table 5.1, the United States, where credit cards have traditionally been popular at the point of sale, shows the highest annual per capita payment card use in 2008. The payment literature stresses consumer payment choice and merchant acceptance in response to price and non-price characteristics. Attempts to determine the main drivers of changes in payment composition across and within countries are difficult due to a lack of time-series data, and often only reported as annual national-level aggregates. Moreover, data on the aggregate cash usage is difficult. Thus, quantifying the pace of migration from paper-based toward electronic means of payment is difficult. To determine the main drivers of change in payment composition across and within countries is difficult because only aggregate data at the national level along (p. 112) with a limited time series dimension is generally available. Given these data problems payment researchers have tried to infer consumer payment behavior from household surveys in Europe and the United States (Stavins, 2001; Hayashi and Klee, 2003; Stix, 2003; Bounie and Francois, 2006; Mester, 2006; Klee, 2008; and Kosse, 2010). Analysis of demographic data indicates that age, education, and income influence the adoption rates of the newer electronic forms of making a payment. However, Schuh and Stavins (2010) show that demographic influences on payment instrument use are often of less importance than the individuals’ assessment of the relative cost, convenience, safety, privacy, and other characteristics of different payment instruments. Other survey-based studies by Jonker (2007) and Borzekowski, Kiser, and Ahmed (2008) that have incorporated similar payment characteristics find that payment instrument characteristics (real or perceived) importantly augment the socio-demographic determinants of the use of electronic payment instruments.8 Page 3 of 21 Digitization of Retail Payments Table 5.1 Annual Per Capita Card Transactions 1988 and 2008 Country 1988 2008 Percent Change Belgium 6.23 87.1 Canada 28.34 187.8 563 France 15.00 102.0 580 Germany 0.76 27.3 3492 Italy 0.33 23.5 7021 Netherlands 0.34 113.7 33,341 Sweden 5.45 176.5 3139 Switzerland 2.34 62.8 2583 United Kingdom 10.47 123.7 1081 United States 36.67 191.1 421 1298 Source: Committee on Payment and Settlement Systems (1993) and (2010). The payment characteristics approach (requiring survey information on which characteristics are favored in one instrument over another) allows estimation of a “price equivalent” trade-off among payment instruments. This approach applies (ordered) probit/logit models to determine price responsiveness relationships among payment instruments. Borzekowski, Kiser, and Ahmed (2008) find a highly elastic response to fees imposed on US PIN debit transactions (an effort by banks to shift users to signature debit cards where bank revenue is higher). Zinman (2009) finds a strong substitution effect between debit and credit cards during 1995–2004, and he concludes that debit card use is more common among consumers who are likely to be credit-constrained in the United States. Another consumer survey suggests that Austrian consumers who often use debit cards hold approximately 20 percent less cash (Stix, 2003). Using French payment data in 2005, Bounie and Francois (2006) estimate the determinants of the probability of (p. 113) a transaction being paid by cash, check, or payment card at the point of sale. They do not only find a clear effect of transaction size but also find evidence that type of good and spending location matter for payment instrument choice. Another approach in the literature has been to infer consumer choice from aggregate data on payment systems and data from industry sources (e.g. Humphrey, Pulley and Vesala, 2000; Garcia-Swartz, Hahn and Layne-Farrar, 2006; Bolt, Humphrey and Uittenbogaard, 2008). Bolt et al. (2008) try to determine the effect of differential transaction-based pricing of payment instruments has on the adoption rate of electronic payments. This is done by comparing the shift to electronic payments during 1990–2004 in two countries—one that has transaction pricing (Norway) and one that does not (the Netherlands). Overall, controlling for country-specific influences, they find that explicit per-transaction payment pricing induces consumers to shift faster to more efficient electronic payment instruments. However, non-price attributes, like convenience and safety, as well as terminal availability may play an even bigger role than payment pricing for payments at the point of sale. There are only few retail payment empirical studies that have used merchant or consumer level transaction data. Klee (2008) provides a simple framework that links payment choice to money holdings. In order to evaluate the model, she uses grocery store scanner data paired with census-tract level demographic information to measure the influence of transaction costs and interest rate sensitivity on payment choice. The data comprise over 10 Page 4 of 21 Digitization of Retail Payments million checkout transactions from September to November 2001. Using a Dubin-McFadden (1984) simultaneous choice econometric specification, Klee finds that a major determinant of consumers’ payment choice is transaction size, with cash being highly favored for small-value transactions. Analysis of the same dataset shows a marked transaction-time advantage for debit cards over checks, helping to explain the increasing popularity of the former. 4. Provision of Payment Services Significant real resources are required to provide payment services. Recent payment cost analyses have shown that the total cost of a nation's retail payment system may easily approach 1 percent of GDP annually (Humphrey, 2010). Even higher cost estimates can be obtained depending on current payment composition and how much of bank branch and ATM network costs are included as being essential for check deposit, cash withdrawal, and card issue and maintenance activity. On the supply side, cost considerations first induced commercial banks to shift cash acquisition by consumers away from branch offices to less costly ATMs. Later, similar cost considerations led banks to try to replace cash, giro and checks with cards using POS terminals although such transactions are likely to lower the (p. 114) demand for ATM cash withdrawals. Greater adoption of mobile payments and online banking solutions will enable a further shift from cash and other paper-based instruments toward more digitized payments. Indeed, payment data —albeit scarcely available—suggest that strong scale economies exist for electronic payments. 4.1. Costs and Benefits of Different Payment Methods Studying the costs to banks to provide payment services is difficult, given the proprietary nature of the cost data. However, there are some European studies that attempt to quantify the real resource costs of several payment services. In these studies, social cost refers to the total cost for society net any monetary transfers between participants reflecting the real resource costs in the production and usage of payment services. For the Netherlands in 2002, Brits and Winder (2005) report that the social costs of all point-of-sale (POS) payments (cash, debit cards, credit cards, and prepaid cards) amounted to 0.65 percent of GDP. The social cost of payment services for Belgium in 2003 was 0.75 percent of GDP (Quaden, 2005). Bergman, Guibourg, and Segendorff (2007) find that the social cost of providing cash, debit card payments, and credit card payments was approximately 0.4 percent of GDP in Sweden for 2002. For Norway, Humphrey, Kim, and Vale (2001) estimate the cost savings from switching from a fully paper-based system (checks and paper “giro,” or a payment in which a payer initiates a transfer from her bank to a payee's bank) to a fully electronic system (debit cards and electronic giro) at the bank level at 0.6 percent of Norway's GDP. Based on a panel of 12 European countries during the period 1987–99, Humphrey et al. (2006) conclude that a complete switch from paper-based payments to electronic payments could generate a total cost benefit close to 1 percent of the 12 nations’ aggregate GDP. These numbers confirm the widespread agreement that the ongoing shift from paper-based payments to electronic payments may result in significant economic gains. Compared with cash, electronic payments also offer benefits in terms of greater security, faster transactions, and better recordkeeping; in addition, electronic payments offer possible access to credit lines.9 Merchants may also benefit from increased sales or cost savings by accepting an array of electronic payment instruments. However, these benefits are often difficult to quantify. Using US retail payments data, Garcia-Swartz, Hahn, and Layne-Farrar (2006) attempt to quantify both the costs and benefits of POS payment instruments.10 They estimate costs and benefits of different components of the payment process and subtract out pure transfers among participants. They find that shifting payments from cash and checks to payment cards improves social welfare as measured by the aggregate surplus to consumers, merchants, and payment providers. However, they also conclude that merchants may pay more for certain electronic payment instruments than some paper-based instruments. (p. 115) 4.2. Economies of Scale and Scope in Payments As more consumers and merchants adopt payment cards, providers of these products may benefit from economies of scale and scope. Size and scalability are important in retail payment systems due to their relatively high capital intensity. In general, electronic payment systems require considerable up-front investments in processing Page 5 of 21 Digitization of Retail Payments infrastructures, require highly secure telecommunication facilities and data storage, and apply complex operational standards and protocols. As a consequence, unit cost should fall as payment volume increases (when appropriately corrected for changes in labor and capital costs). In addition, scope economies come into play when different payment services can be supplied on the same electronic network in a more cost-efficient way than the “stand-alone” costs of providing these services separately. In the United States, being able to operate on a national level allowed some issuers (banks that issue cards to consumers), acquirers (banks that convert payment card receipts into bank deposits for merchants), and payment processors to benefit from economies of scale and scope. We discuss two large consolidations that occurred within the Federal Reserve System over the last two decades that resulted in large cost savings. First, the Federal Reserve's real-time gross settlement large-value payment system, Fedwire, consolidated its 12 separate payment processing sites into a single site in 1996. As a result, Fedwire's average cost per transaction fell by about 62 percent in real terms from scale economies due to the expanded volume and technological change which lowered processing and telecommunication costs directly.11 A similar process occurred in Europe where in 2007, 15 separate national real time gross settlement systems were consolidated into one single technical platform, TARGET2, that guaranteed a harmonized level of service for European banks combined with one single transaction price for domestic and cross-border payments. Second, with the passage of Check Clearing for the 21st Century Act in 2003 and the reduction in the number of checks written in the United States, the Federal Reserve reduced the number of its check processing sites from 45 to 1 by March 2010. Today, almost 99 percent of checks are processed as images, thus enabling greater centralization in check processing. In addition, many checks are converted to ACH payments. Both check imaging and conversion have resulted in significant cost savings to the Federal Reserve and market participants. Some European payment providers might enjoy similar scale and scope benefits in the future as greater crossborder harmonization occurs with the introduction of the Single Euro Payments Area (SEPA).12 The goal of SEPA, promoted by the European Commission, is to facilitate the emergence of a competitive, intra-European goods market by making cross-border payments as easy as domestic transactions. Separate domestic national payments infrastructures are to be replaced with a pan-European structure that would lower payment costs through economies of scope and scale. Volume expansion can best be achieved by consolidating processing operations across European borders. (p. 116) One of the first European scale economies study on payment systems was carried out by Khiaonarong (2003). He estimates a simple loglinear cost function by using data of 21 payment systems and finds substantial scale economies.13 In Bolt and Humphrey (2007), a data set including 11 European countries over 18 years is used to explain movements of operating costs in the banking sector as a function of transaction volumes of four separate payment and delivery instruments (card payments, bill payments, ATMs, and branch offices), controlling for wages and capital costs. Their primary focus is on scale economies of card payments. In particular, using a translog function specification, the average scale economy is (significantly) estimated in the range 0.25–0.30, meaning that a doubling of payment volume corresponds to only a 25 to 30 percent increase in total costs. Consequently, volume expansion should lead to significantly lower average costs per transaction. Based on cost data specific to eight European payment processor operations over the period 1990 to 2005, Beijnen and Bolt (2009) obtain similar estimates of payment scale economies, which allow them to quantify the potential benefits of SEPA arising from consolidation of electronic payment processing centers across the euro area. Finally, Bolt and Humphrey (2009) estimate payment scale and scope economies using previously unavailable (confidential) individual bank data for the Netherlands from 1997 to 2005. Their analysis confirms the existence of strong payment scale economies, thus furthering the goal of SEPA. One key result stands out: payment costs can be markedly reduced through consolidation of payment processing operations to realize economies of scale. Ultimately, this allows banks, consumers, and merchants to benefit from these cost efficiencies in the form of lower payment fees. However, how each participant benefits from this reduction in payment costs and exactly how it is allocated in terms of lower payment and service fees, lower loan rates, higher deposit rates, or higher bank profit is an issue of great interest to public authorities. 5. Economic Surplus and Cross Subsidies Page 6 of 21 Digitization of Retail Payments To study the optimal structure of fees between consumers and merchants in payment markets, economists have developed the two-sided market or platform framework. This literature combines the multiproduct firm literature, which studies how firms set prices on more than one product, with the network economics literature, which studies how consumers benefit from increased participation of consumers in the network.14 The price structure or balance is the share of the total price of the payment service that each type of end-user pays. Rochet and Tirole (2006b) define a two-sided market as a market where end-users are unable to negotiate prices (p. 117) among themselves and the price structure affects the total volume of transactions.15 In practice, the existence of nosurcharge rules do not allow consumers and merchants to negotiate prices based on the underlying costs of the payment instrument used. Furthermore, even in jurisdictions where such practices have been outlawed, most merchants have been reluctant to differentiate their prices. An important empirical observation of two-sided markets is that platforms tend to heavily skew the price structure to one side of the market to get both sides “on board,” using one side as a “profit center” and the other side as a “loss leader,” or at best financially neutral.16 In the rest of this section, we will discuss several externalities that arise in payment networks.17 5.1. Adoption and Usage Externalities A key externality examined in the payment card literature is the ability of the network to convince both consumers and merchants to participate in a network. Baxter (1983) argues that the equilibrium quantity of payment card transactions occurs when the total transactional demand for payment card services, which are determined by consumer and merchant demands jointly, is equal to the total transactional cost for payment card services, including both issuer and acquirer costs.18 A consumer's willingness to pay is based on her net benefits received. The consumer will participate if her net benefit is greater than or equal to the fee. Similarly, if the merchants’ fee is less than or equal to the net benefit they receive, merchants will accept cards. Net benefits for consumers and merchants are defined by the difference in benefits from using or accepting a payment card and using or accepting an alternative payment instrument. Pricing each side of the market based on marginal cost—as would be suggested by economic theory for one-sided competitive markets—need not yield the socially optimal allocation. To arrive at the socially optimal equilibrium, a side payment may be required between the issuer and acquirer if there are asymmetries of demand between consumers and merchants, differences in costs to service consumers and merchants, or both. This result is critically dependent on the inability of merchants to price differentiate between card users and those who do not use cards or among different types of card users. While most economists and antitrust authorities agree that an interchange fee may be necessary to balance the demands of consumers and merchants resulting in higher social welfare, the level of the fee remains a subject of debate. Schmalensee (2002) extends Baxter's (1983) analysis by considering issuers and acquirers that have market power, but still assumes that merchants operate in competitive markets. His results support Baxter's conclusions that the interchange fee balances the demands for payment services by each end-user type and the cost to banks to provide them. Schmalensee finds that the profit-maximizing interchange fee of issuers and acquirers may also be socially optimal. (p. 118) 5.2. Instrument-Contingent Pricing In many jurisdictions, merchants are not allowed to add a surcharge for payment card transactions because of legal or contractual restrictions. No-surcharge restrictions do not allow merchants to impose surcharges for payment card purchases. However, merchants may be allowed to offer discounts for noncard payments instead of surcharges.19 If consumers and merchants were able to negotiate prices based on differences in costs that merchants face and the benefits that both consumers and merchants receive, the interchange fee would be neutral, assuming full pass-through. The interchange fee is said to be neutral if a change in the interchange fee does not change the quantity of consumer purchases and the profit level of merchants and banks. There is general consensus in the payment card literature that if merchants were able to recover costs to accept a given payment instrument directly from those consumers that use it, the impact of the interchange fee would be severely dampened. Even if price differentiation based on the payment instrument used is not common, the possibility to do so may enhance the merchants’ bargaining power in negotiating their fees.20 Merchants can exert downward pressure on fees by having the possibility to set instrument-contingent pricing. Payment networks may prefer non-instrument- Page 7 of 21 Digitization of Retail Payments contingent pricing because some consumers may not choose payment cards if they had to explicitly pay for using them at the point of sale (POS). Carlton and Frankel (1995) extend Baxter (1983) by considering when merchants are able to fully pass on payment processing costs via higher consumption goods prices. They find that an interchange fee is not necessary to internalize the externality if merchants set pricing for consumption goods based on the type of payment instrument used. Furthermore, they argue that cash users are harmed when merchants set one price because they subsidize card usage.21 Schwartz and Vincent (2006) study the distributional effects among cash and card users with and without nosurcharge restrictions. They find that the absence of pricing based on the payment instrument used increases network profit and harms cash users and merchants.22 The payment network prefers to limit the merchant's ability to separate card and cash users by forcing merchants to charge a uniform price to all of its customers. Issuer rebates to card users boost their demand for cards while simultaneously forcing merchants to absorb part of the corresponding rise in the merchant fee, because any resulting increase in the uniform good's price must apply equally to cash users. Gans and King (2003) argue that, as long as there is “payment separation,” the interchange fee is neutral regardless of the market power of merchants, issuers, and acquirers. When surcharging is costless, merchants will implement pricing based on the payment instrument used, taking away the potential for cross-subsidization across payment instruments and removing the interchange fee's role in balancing the demands of consumers and merchants. In effect, the cost pass-through is such that lower consumer card fees (due to higher interchange fees) are exactly offset by higher goods prices from merchants. Payment separation can occur if one of (p. 119) the following is satisfied: There are competitive merchants, and they separate into cash-accepting or card-accepting categories, in which each merchant only serves one type of customer and is prevented from charging different prices; or merchants are able to fully separate customers who use cash from those who use cards by charging different prices. 5.3. Merchant, Network, and Issuer Competition When asking merchants why they accept certain types of payment cards if they are too costly, they answer that they would lose business to their competitors. Rochet and Tirole (2002) were the first to consider business stealing as a motivation for merchants to accept payment cards. Rochet and Tirole study the cooperative determination of the interchange fee by member banks of a payment card association in a model of two-sided markets with network externalities. They develop a framework in which banks and merchants may have market power and consumers and merchants decide rationally on whether to use or accept a payment card. In particular, Rochet and Tirole consider two identical Hotelling merchants in terms of their net benefits of accepting a payment cards. Consumers face the same fixed fee but are heterogeneous in terms of the net benefits they derive from using the payment card. They assume that the total number of transactions is fixed and changes in payment fees do not affect the demand for consumption goods. They have two main results. First, the interchange fee that maximizes profit for the issuers may be greater than or equal to the socially optimal interchange fee, depending on the issuers’ margins and the cardholders’ surplus. An interchange fee set too high may lead to overprovison of payment card services. Second, merchants are willing to pay more than the socially optimal fee if they can steal customers from their competitors. Payment card networks can exploit each merchant's eagerness to obtain a competitive edge over other merchants. Remarkably, this rent extraction has also some social benefits since, on the consumer side, it offsets the underprovision of cards by issuers with market power. However, overall social welfare does not improve when merchants steal customers from their competitors by accepting payment cards. Wright (2004) extends Rochet and Tirole (2002) by considering a continuum of industries where merchants in different industries receive different benefits from accepting cards. In his environment, consumers and merchants pay per transaction fees. Each consumer buys goods from each industry. Issuers and acquirers operate in markets with imperfect competition. He assumes that consumers face the same price regardless of which instrument they use to make the purchase. Similar to Rochet and Tirole (2002), Wright concludes that the interchange fee that maximizes overall social welfare is generally higher than the interchange fee that maximizes the number of Page 8 of 21 Digitization of Retail Payments transactions. Economic theory suggests that competition among suppliers of goods and services generally reduces prices, increases output, and improves welfare. However, (p. 120) within two-sided market framework, network competition may yield an inefficient price structure. A key aspect of network competition is the ability of end-users to participate in more than one network. When end-users participate in more than one network, they are said to be “multihoming.” If they connect only to one network, they are said to be “single-homing.” As a general finding, competing networks try to attract end-users who tend to single-home, since attracting them determines which network has the greater volume of business. Using data from Visa, Rysman (2007) demonstrates that even though consumers carry multiple payment cards in their wallet, they tend to use the same card for most of their purchases.23 Accordingly, the price structure is tilted in favor of end-users who single-home.24 Some models of network competition assume that the sum of consumer and merchant fees is constant and focus on the price structure. Rochet and Tirole (2003) find that the price structures for a monopoly network and competing platforms may be the same, and if the sellers’ demand is linear, this price structure in the two environments generates the highest welfare under a balanced budget condition. Guthrie and Wright (2007) extend Rochet and Tirole (2003) by assuming that consumers are able to hold one or both payment cards. They find that network competition can result in higher interchange fees than those that would be socially optimal. In their model, Guthrie and Wright take into account that in a payment network, merchants, who are on one side of the market, compete to attract consumers, who are on the other side. This asymmetry causes payment system competition to over-represent interests of cardholders because they generally single-home causing heterogenous merchants to be charged more and cardholders less. This skewed pricing effect is reinforced when consumers are at least as important as merchants in determining which card will be adopted by both sides. The result that system competition may increase interchange fees illustrates the danger of using onesided logic to make inferences in two-sided markets. Chakravorti and Roson (2006) consider the effects of network competition on total price and on price structure where networks offer differentiated products. They only allow consumers to participate in one card network, whereas merchants may choose to participate in more than one network. However, unlike Guthrie and Wright (2007) and Rochet and Tirole (2003), Chakravorti and Roson consider fixed fees for consumers. They compare welfare properties when the two networks operate as competitors and as a cartel, where each network retains demand for its products from end-users but the networks set fees jointly. Like Rochet and Tirole (2003) and Guthrie and Wright (2007), Chakravorti and Roson (2006) find that competition does not necessarily improve or worsen the balance of consumer and merchant fees from the socially optimal one. However, they find that the welfare gain from the drop in the sum of the fees from competition is generally larger than the potential decrease in welfare from less efficient fee structures. Competition does not necessarily improve the balance of prices for two-sided markets. Furthermore, if competition for cardholders is more intense because (p. 121) consumers ultimately choose the payment instrument, issuers may provide greater incentives to attract consumers even if both issuers belong to the same network. If issuers have greater bargaining power to raise interchange fees, they can use this power to partially offset the cost of consumer incentives.25 5.4. Credit Functionality of Payment Cards The payment card literature has largely ignored the benefits of consumer credit.26 Given the high level of antitrust scrutiny targeted toward credit card fees, including interchange fees, this omission in most of the academic literature is rather surprising. In the long run, aggregate consumption over consumers’ lives may not differ because of access to credit, but such access may enable consumers to increase their utility. In addition to extracting surplus from all consumers and merchants, banks may extract surplus from consumers that borrow in the form of finance charges.27 Chakravorti and Emmons (2003) consider the costs and benefits of consumer credit when consumers are subject to income shocks after making their credit card purchases and some are unable to pay their credit card debt. Chakravorti and Emmons assume that all markets for goods and payment services are competitive. They impose a Page 9 of 21 Digitization of Retail Payments participation constraint on individuals without liquidity constraints such that the individuals will only use cards if they are guaranteed the same level of consumption as when they use cash including the loss of consumption associated with higher prices for consumption goods. To our knowledge, they are the first to consider the credit payment functionality of credit cards. Observing that over 75 percent of US card issuer revenue is derived from cash-constrained consumers, they consider the viability of the credit card system if it were completely funded by these types of consumers.28 They find that if consumers sufficiently discount future consumption, liquidity-constrained consumers who do not default would be willing to pay all credit card network costs ex ante, resulting in all consumers being better off. However, they also find that the inability of merchants to impose instrument-contingent prices results in a lower level of social welfare because costly credit card infrastructure is used for transactions that do not require credit extensions. Chakravorti and To (2007) consider an environment with a monopolist bank that serves both consumers and merchants where the merchants absorb all credit and payment costs in a two-period dynamic model. Chakravorti and To depart from the payment card literature in the following ways. First, rather than taking a reduced-form approach where the costs and benefits of payment cards are exogenously assigned functional forms, they construct a model that endogenously yields costs and benefits to consumers, merchants, and banks from credit card use. Second, their model considers a dynamic setting where there are intertemporal tradeoffs for all participants. Third, they consider consumption and income uncertainty. (p. 122) Their model yields the following results. First, the merchants’ willingness to pay bank fees increases as the number of credit card consumers without income increases. Note that up to a point, merchants are willing to subsidize credit losses in exchange for additional sales. Second, a prisoner's dilemma situation may arise: Each merchant chooses to accept credit cards, but by doing so, each merchant's discounted two-period profit is lower. Unlike other models, business stealing occurs across time and across monopolist merchants. 5.5. Competition Among Payment Instruments Most of the payment card literature ignores competition between payment instruments.29 If consumers carry multiple types of payment instruments, merchants may be able to steer them away from more costly payment instruments.30 Rochet and Tirole (2011) argue that merchants may choose to decline cards after they have agreed to accept them. In their model a monopoly card network would always select an interchange fee that exceeds the level that maximizes consumer surplus. If regulators only care about consumer surplus, a conservative regulatory approach is to cap interchange fees based on retailers’ net avoided costs from not having to provide credit themselves. This always raises consumer surplus compared to the unregulated outcome, sometimes to the point of maximizing consumer surplus. This regulatory cap is conceptually the same as the “tourist test” where the merchant accepts cards even when it can “effectively steer” the consumer to use another payment instrument. However, if the consumer is unable to access cash or another form of payment, the merchant would lose the sale. Merchants may steer consumers through price incentives, if allowed to do so. Bolt and Chakravorti (2008a) study the ability of banks and merchants to influence the consumers’ choice of payment instrument when they have access to three payment forms—cash, debit card, and credit card. In their model, consumers only derive utility from consuming goods from the merchant they are matched to in the morning. Merchants differ on the types of payment instruments that they accept and type of consumption good they sell. Each merchant chooses which instruments to accept based on its production costs and merchant heterogeneity is based on these differences in production costs. They consider the merchants’ ability to pass on payment processing costs to consumers in the form of higher uniform and differentiated goods prices. Unlike most two-sided market models, where benefits are exogenous, they explicitly consider how consumers’ utility and merchants’ profits increase from additional sales resulting from greater security and access to credit.31 Bolt and Chakravorti's (2008a) key results can be summarized as follows. First, with sufficiently low processing costs relative to theft and default risk, the social planner sets the credit card merchant fee to zero, completely internalizing the card acceptance externality. Complete merchant acceptance maximizes card usage at (p. 123) the expense of inefficient cash payments.32 The bank may also set the merchant fees to zero, but only if Page 10 of 21 Digitization of Retail Payments merchants are able to sufficiently pass on their payment fees to their consumers or if their payment processing costs are zero. Second, if the real resource cost of payment cards is sufficiently high, the social planner sets a higher merchant fee than the bank does, resulting in lower card acceptance and higher cash usage. Third, bank profit is higher when merchants are unable to pass on payment costs to consumers because the bank is better able to extract merchant surplus. On the other hand, full pass-through would retrieve the neutrality result of Gans and King (2003) where all payment costs are shifted onto consumers’ shoulders relieving potential two-sided tension. However, if merchants need to absorb part of these cost as well, two-sided externalities remain to play a role for optimal payment pricing. Finally, in their model, the relative cost of providing debit and credit cards determines whether the bank will provide both or only one type of payment card. 6. Market Interventions In this section, we discuss several market interventions in various jurisdictions. Specifically, we focus on three different types of market interventions. First, we discuss the removal of pricing restrictions placed on merchants that prevent them from surcharging customers that are using certain types of payment instruments. Second, we study the impact of adoption and usage of payment cards when public authorities impose caps on interchange fees. Third, we discuss the forced acceptance of all types of cards, that is, credit, debit, and prepaid, when merchants enter into contracts with acquirers. 6.1. Removal of No-Surcharge Policies There are several jurisdictions where merchants are able to surcharge card transactions. Most of the academic research suggests that if merchants are allowed to surcharge, the level of the interchange fee would be neutral. If the interchange fee is neutral, regulating the interchange fee would have little impact. In this section, we explore whether merchants surcharge if they are allowed to do so. To encourage better price signals, the RBA removed no-surcharge restrictions in 2002. The Australian authorities argued that consumers did not receive the proper price incentives to use debit cards, the less costly payment instrument. The Reserve Bank of Australia (RBA) reported that the average cost of the payment functionality of the credit card was AUS$0.35 higher than a debit card using a consistent AUS$50 transaction size.33 (p. 124) While most Australian merchants do not impose surcharges for any type of payment card transaction today, the number of merchants surcharging credit card transactions continues to increase. At the end of 2007, around 23 percent of very large merchants and around ten percent of small and very small merchants imposed surcharges. The average surcharge for MasterCard and Visa transactions is around one percent, and that for American Express and Diners Club transactions is around two percent (Reserve Bank of Australia, 2008a). Using confidential data, the Reserve Bank of Australia (2008a) also found that if one network's card was surcharged more than another networks’ cards, consumers dramatically reduced their use of the card with the surcharge. Differential surcharging based on which network the card belongs to may result in greater convergence in merchant fees across payment card networks. In the United States, merchants are allowed to offer cash discounts but may not be allowed to surcharge credit card transactions. In the 1980s, many US gas stations explicitly posted cash and credit card prices. Barron, Staten, and Umbeck (1992) report that gas station operators imposed contingent-instrument pricing when their credit card processing costs were high but later abandoned this practice when acceptance costs decreased because of new technologies such as electronic terminals at the point of sale. Recently, some gas stations brought back price differentiation based on payment instrument type, citing the rapid rise in gas prices and declining profit margins. On the other hand, in some instances, policymakers may prefer if merchants did not surcharge certain types of transactions. For example, Bolt, Jonker, and van Renselaar (2010) find that a significant number of merchants surcharge debit transactions vis-à-vis cash in the Netherlands. Debit card surcharges are widely assessed when purchases are below 10 euro, suggesting that merchants are unwilling to pay the fixed transaction fee below this threshold. They find that merchants may surcharge up to four times their fee. In addition, when these surcharges are removed, they argue that consumers start using their debit cards for these small payments, suggesting that merchant price incentives do affect consumer payment choice. Interestingly, in an effort to promote a more efficient payment system, the Dutch central bank has supported a public campaign to encourage retailers to stop Page 11 of 21 Digitization of Retail Payments surcharging to encourage consumers to use their debit cards for small transactions. This strategy appears to be successful. In 2009, debit card payments below ten euro accounted for more than 50 percent of the total annual growth of almost 11 percent in debit card volume. There are instances when card payments are discounted vis-à-vis cash payments. The Illinois Tollway charges motorists who use cash to pay tolls twice as much as those who use toll tags (called I-PASS), which may be loaded automatically with credit and debit cards when the level of remaining funds falls below a certain level (Amromin, Jankowski, and Porter, 2007). In addition to reducing cash handling costs, the widespread implementation of toll tags decreased not only congestions at toll booths but also pollution from idling vehicles waiting to pay tolls, since tolls could be collected as cars drove at highway speeds through certain points on the Illinois Tollway. (p. 125) The ability for merchants to charge different prices is a powerful incentive to convince consumers to use a certain payment instrument. In reality, merchants may surcharge or discount card transactions depending on their underlying cost structures along with benefits accrued. However, in some instances, surcharges may result in a less desirable outcome as evidenced in the Dutch example suggesting a potential holdup problem whereby merchants impose higher surcharges than their costs. Furthermore, there is also evidence of cash surcharges suggesting that card acceptance costs are lower than the costs of handling cash 6.2. Regulation of Interchange Fees There are several jurisdictions where interchange fees were directly regulated or significant pressure was exerted by the public authorities on networks to reduce their interchange fees.34 In this section, we will discuss the impact of interventions in four jurisdictions—Australia, Spain, the European Union, and the United States. 6.2.1. Australia In 2002, the Reserve Bank of Australia (RBA) imposed weighted-average MasterCard and Visa credit card interchange fee caps and later imposed per transaction targets for debit cards. As of April 2008, the weightedaverage credit card inter-change fees in the MasterCard and Visa networks must not exceed 0.50 percent of the value of transactions. The Visa debit weighted-average interchange fee cap must not exceed 12 cents (Australian) per transaction. The EFTPOS (electronic funds transfer at point of sale) interchange fees for transactions that do not include a cash-out component must be between four cents (Australian) and five cents (Australian) per transaction. The Reserve Bank of Australia (2008a) reports that the interchange fee regulation, coupled with the removal of the no-surcharge rule, improved the price signals that consumers face when deciding which payment instruments to use. Specifically, annual fees for credit cards increased and the value of the rewards decreased. The Reserve Bank of Australia (2008a) calculates that for an AUS$100 transaction, the cost to consumers increased from – AUS$1.30 to –AUS$1.10 for consumers who pay off their balances in full every month. A negative per transaction cost results when card benefits such as rewards and interest-free loans are greater than payment card fees.35 Those who oppose the Australian interchange fee regulation argue that consumers have been harmed by reduced rewards and higher fees and have not shared in the cost savings—in terms of lower prices for goods and services. However, measuring price effects over time of interchange fee regulation is difficult.36 6.2.2. Spain Unlike in Australia, the Ministry of the Economy, the Ministry of Industry, Tourism, and Trade along with the antitrust authority, and not the central bank, intervened in payment card markets in Spain several times during the period 1999 to (p. 126) 2009. Part of the motivation was based on directives by the European Commission regarding fees that were set by networks that had significant market power. These regulations had significant impact on debit and credit card usage. Over the period 1997–2007, debit card transactions increased from 156 million to 863 million and credit card transactions increased from 138 million to 1.037 billion. Carbó-Valverde, Chakravorti, and Rodriguez Fernandez (2009) study the effects of interchange fee reductions in Spain from 1997 to 2007. To our knowledge, they are the first to use bank-level data to study the impact of several episodes of interchange fee reductions for debit and credit cards resulting from moral suasion and agreements Page 12 of 21 Digitization of Retail Payments between market participants intermediated by the government authorities. They demonstrate that merchants benefited from lower fees and consumers benefited from greater merchant acceptance. Surprisingly, they found that issuer revenues increased during the period when interchange fees decreased. While the effect of these reductions is positive on banks’ revenues, their effect on banks’ profits could not be determined because of data limitations. Furthermore, there may be a critical interchange fee below which issuer revenue decreases. 6.2.3. European Commission In December 2007, the European Commission (EC) ruled that the multilateral interchange fees for cross-border payments in the European Union applied by MasterCard Europe violated Council Regulation (EC) No. 1/2003. The EC argued that MasterCard's fee structure restricted competition among acquiring banks and inflated the cost of card acceptance by retailers without leading to proven efficiencies.37 In response, MasterCard reached an interim understanding with the European Commission on these interchange fees for cross-border consumer payments in the EU in April 2009. Effective July 1, 2009 MasterCard, Europe has established interchange fees for consumer card transactions that, on average, will not exceed 30 basis points for credit cards and 20 basis points for debit cards. With these fee changes, the EC will not further pursue MasterCard either for non-compliance with its December 2007 decision or for infringing the antitrust rules. The EC conducted a separate antitrust investigation against Visa and will monitor the behavior of other market players to ensure that competition is effective in this market to the benefit of merchants and consumers. The EC and Visa have agreed to 20 basis point debit card interchange fees but have not agreed to the level of credit card interchange fees. The dialogue between Visa and MasterCard vis-à-vis the Commission has to date not led to an agreement concerning the application of the “merchant indifference methodology” based on the tourist test to consumer credit (and deferred debit) transactions—discussions on this issue continue. 6.2.4. United States As part of the financial reform bill signed into law on July 21, 2010, a section of Title 10 of the Dodd-Frank Wall Street Reform and Consumer Protection Act grants the Federal Reserve Board the authority to set rules regarding the setting of debit card (p. 127) interchange fees that are “reasonable and proportional to cost.” Financial institutions with less than $10 billion in assets are exempt. Debit cards include payment cards that access accounts at financial institutions to make payment along with prepaid cards. Certain types of prepaid cards are exempt such as those disbursing funds as a part of government-sponsored programs or geared toward the underbanked or lower-income households.38 The level of the debit card interchange fee has not been decided at the time of writing. 6.3. Honor-All-Cards Rules A payment card network may require merchants that accept one of its payment products to accept all of its products.39 Such a rule is a type of honor-all-cards rule. In other words, if a merchant accepts a network's credit card, it must accept all debit and prepaid cards from that network. In the United States, around 5 million merchants sued the two major networks, MasterCard and Visa, over the required acceptance of the network's signature-based debit card when accepting the same network's credit card.40 The case was settled out of court. In addition to a monetary settlement, MasterCard and Visa agreed to decouple merchants’ acceptance of their debit and credit products. While few merchants have declined one type of card and accepted another type, the decoupling of debit and credit card acceptance may have increased bargaining power for merchants in negotiating fees. As part of the payment system reforms in Australia, MasterCard and Visa were mandated to decouple merchants’ acceptance of their debit and credit cards as well. The Payments System Board (Reserve Bank of Australia, 2008b) is unaware of any merchant that continues to accept debit cards but does not accept credit cards from the same network. 7. Conclusion Technological advances in mobile phone technology have the potential to replace many remaining paper-based transactions. This will also increase the usage of electronic payment usage. In other words, how rapidly payment Page 13 of 21 Digitization of Retail Payments innovations are introduce and adopted is critically dependent on the potential profitability of the retail payment system as a whole. However, the rate at which these shifts will occur depends on the underlying benefits and costs to payment system participants. Most policymakers and economists agree that the digitization of payments is socially beneficial. However, there is considerable debate regarding the optimal pricing of these payment services. Payment markets are complex with many participants engaging in a series of interrelated bilateral transactions. (p. 128) The determination of optimal prices is difficult for several reasons. First, there are significant scale and scope economies in payment processing because of large fixed costs to setup sophisticated secure networks to process, clear and settle payment transactions. Thus, established payment providers may generally enjoy some level of market power because these markets are generally not contestable. Second, payment networks must convince two distinct sets of end users—consumers and merchants—to simultaneously participate. Networks often set asymmetric prices to get both sides on board. Such pricing is based on cost to serve end-users, as well as their demand elasticities. It is extremely difficult for policymakers to disentangle optimal pricing strategies from excessive rent extraction. Third, efficiency of payment systems is measured not only by the costs of resources used, but also by the social benefits generated by them. Measuring individual and social benefits is particularly difficult. The central question is whether the specific circumstances of payment markets are such that intervention by public authorities can be expected to improve economic welfare. The theoretical literature on payment cards continues to grow. However, there are a few areas of payment economics that deserve greater attention. First, what is the effect of the reduction of banks’ and networks’ surplus extraction on future innovation? Second, how should payment system participants pay for fraud containment and distribute the losses when fraud occurs? Third, should public entities step in and start providing payment services and, if so, at what price? Finally, empirical studies about payment system pricing using data from payment networks and providers are extremely scarce. Such analysis would be helpful in understanding how effective regulatory interventions were in meeting the stated objectives and studying any potential unintended consequences. We hope that recent regulatory changes in different parts of the world will generate rich sets of data that can be exploited by economists to test how well the theories fit the data. References Adams, R., Bauer, P., Sickles R. 2004. Scale Economies, Scope Economies, and Technical Change in Federal Reserve Payment Processing. Journal of Money, Credit and Banking 36(5), pp. 943–958. Agarwal, S., Chakravorti, S., Lunn, A., 2010. Why Do Banks Reward Their Customers to Use Their Credit Cards, Federal Reserve Bank of Chicago Working Paper, WP-2010–19. Alvarez, F., Lippi, F., 2009. Financial Innovation and the Transactions Demand for Cash, Econometrica, 77 (2), pp. 363–402. Amromin, G., Chakravorti, S., 2009. Whither Loose Change? The Diminishing Demand for Small Denomination Currency. Journal of Money, Credit and Banking 41(2–3), pp. 315–335. Amromin, G., Jankowski, C., Porter, R., 2007. Transforming Payment Choices by Doubling Fees on the Illinois Tollway. Economic Perspectives, Federal Reserve Bank of Chicago 31(2), pp. 22–47. Armstrong, M., 2006. Competition in Two-sided Markets. RAND Journal of Economics 37(3), pp. 668–691. Ausubel, L., 1991. The Failure of Competition in the Credit Card Market. American Economic Review 81(1), pp. 50– 81. Barron, J., Staten, M., Umbeck, J., 1992. Discounts for Cash in Retail Gasoline Marketing. Contemporary Policy Issues 10(4), pp. 89–102. Page 14 of 21 Digitization of Retail Payments Baxter, W., 1983. Bank Interchange of Transactional Paper: Legal and Economic Perspectives. Journal of Law and Economics 26(3), pp. 541–588. Bedre-Defolie, Ö., Calvano, E., 2010. Pricing Payment Cards. Toulouse School of Economics and Princeton University, mimeo. Beijnen, C., Bolt, W., 2009. Size Matters: Economies of Scale in European Payments Processing. Journal of Banking and Finance 33(2), pp. 203–210. Benner, K., 2008. Visa's Record IPO Rings Up 28 Percent Gain. CNNMoney.com, March 19. Available at: http://money.cnn.com/2008/03/19/news/companies/visa_ipo_opens.fortune/index.htm Bergman, M., Guibourg, G., Segendorff, B., 2007. The Costs of Paying—Private and Social Costs of Cash and Card. Sveriges Riksbank, working paper, No. 212. Bolt, W., 2008. The European Commission's Ruling in MasterCard: A Wise Decision? GCP, April 1. Available at: www.globalcompetitionpolicy.org/index.php?id=981&action=907. Bolt, W., Carbó-Valverde, S., Chakravorti, S., Gorjón, S., Rodríguez Fernández, F., 2010. What is the Role of Public Authorities in Retail Payment Systems? Federal Reserve Bank of Chicago Fed Letter, 280a, pp. 1–4. Bolt, W., Chakravorti, S., 2008a. Consumer Choice and Merchant Acceptance of Payment Media. Federal Reserve Bank of Chicago Working Paper, WP-2008–11. Bolt, W., Chakravorti, S., 2008b. Economics of Payment Cards: A Status Report. Economic Perspectives, Federal Reserve Bank of Chicago 32(4), pp. 15–27. Bolt, W., Humphrey, D., 2007. Payment Network Scale Economies, SEPA, and Cash Replacement. Review of Network Economics 6(4), pp. 453–473. Bolt, W., Humphrey, D., 2009. Payment Scale Economies from Individual Bank Data. Economics Letters 105(3), pp. 293–295. (p. 132) Bolt, W., Humphrey, D., Uittenbogaard, R., 2008. Transaction Pricing and the Adoption of Electronic Payments: A Cross-country Comparison. International Journal of Central Banking 4(1), 89–123. Bolt, W., Jonker, N., van Renselaar, C., 2010. Incentives at the Counter: An Empirical Analysis of Surcharging Card Payments and Payment Behavior in the Netherlands. Journal of Banking and Finance 34(8), pp. 1738–1744. Bolt, W., Schmiedel, H., 2011. Pricing of Payment Cards, Competition, and Efficiency: A Possible Guide for SEPA. Annals of Finance, pp. 1–21. Bolt, W., Tieman, A., 2008. Heavily Skewed Pricing in Two-Sided Markets. International Journal of Industrial Organization 26(5), pp. 1250–1255. Borzekowski, R., Kiser, E., Ahmed, S., 2008. Consumers’ Use of Debit Cards: Patterns, Preferences, and Price Response. Journal of Money, Credit and Banking 40(1), pp. 149–172. Bounie, D., Francois, A., 2006. Cash, Check or Bank Card? The Effects of Transaction Characteristics on the Use of Payment Instruments. Telecom Paris Tech, working paper ESS-06–05. Bourreau, M., Verdier, M., 2010. Private Cards and the Bypass of Payment Systems by Merchants. Journal of Banking and Finance 34(8), pp. 1798–1807. Bradford, T., Hayashi, F., 2008. Developments in Interchange Fees in the United States and Abroad. Payments System Research Briefing, Federal Reserve Bank of Kansas City, April. Brito, D., Hartley, P., 1995. Consumer Rationality and Credit Cards. Journal of Political Economy 103(2), pp. 400– 433. Brits, H., Winder, C., 2005. Payments Are no Free Lunch. Occasional Studies, De Nederlandsche Bank 3(2), pp. 1– Page 15 of 21 Digitization of Retail Payments 44. Carbó-Valverde, S., Chakravorti, S., Rodriquez Fernandez, F., 2009. Regulating Two-sided Markets: An Empirical Investigation. Federal Reserve Bank of Chicago Working Paper, WP-2009–11. Carbó-Valverde, S., Humphrey, D., Liñares Zegarra, J.M., Rodríguez Fernández, F., 2008. A Cost–Benefit Analysis of a Two-sided Card Market. Fundación de las Cajas de Ahorros (FUNCAS), working paper, No. 383. Carbó-Valverde, S., Liñares Zegarra, J.M., 2009. How Effective Are Rewards Programs in Promoting Payment Card Usage? Empirical Evidence. Fundación BBVA, working paper, No. 1. Carlton, D., Frankel, A., 1995. The Antitrust Economics of Credit Card Networks. Antitrust Law Journal, 63(2), pp. 643–668. Chakravorti, S., 2007. Linkages Between Consumer Payments and Credit. In: S. Agarwal, Ambrose, B.W. (Eds.), Household Credit Usage: Personal Debt and Mortgages, New York, Palgrave MacMillan, pp. 161–174. Chakravorti, S., 2010. Externalities in Payment Card Networks: Theory and Evidence. Review of Network Economics, 9(2), pp. 99–134. Chakravorti, S., Emmons, W., 2003. Who Pays for Credit Cards? Journal of Consumer Affairs, 37(2), pp. 208–230. Chakravorti, S., Lubasi, V., 2006. Payment Instrument Choice: The Case of Prepaid Cards. Economic Perspectives, Federal Reserve Bank of Chicago 30(2), pp. 29–43. Chakravorti, S., Roson, R., 2006. Platform Competition in Two-sided Markets: The Case of Payment Networks. Review of Network Economics 5(1), pp. 118–143. Chakravorti, S., Shah, A., 2003. Underlying Incentives in Credit Card Networks. Antitrust Bulletin 48(1), pp. 53–75. (p. 133) Chakravorti, S., To, T., 2007. A Theory of Credit Cards. International Journal of Industrial Organization 25(3), pp. 583–595. Chang, H., Evans, D., Garcia Swartz, D., 2005. The Effect of Regulatory Intervention in Two-sided Markets: An Assessment of Interchange-Fee Capping in Australia. Review of Network Economics 4(4), pp. 328–358. Ching, A., Hayashi, F., 2010. Payment Card Rewards Programs and Consumer Payment Choice. Journal of Banking and Finance 34(8), pp. 1773–1787. Committee on Payment and Settlement Systems 1993 and 2010. Statistics on Payment and Settlement Systems in Selected Countries, Basel, Switzerland: Bank for International Settlements. Constantine, L., 2009. Priceless, New York: Kaplan Publishing. Donze, J., Dubec, I., 2009. Pay for ATM Usage: Good for Consumers, Bad for Banks? Journal of Industrial Economics, 57(3), pp. 583–612. Dubin, J., McFadden, D., 1984. An Econometric Analysis of Residential Electric Appliance Holdings and Consumption. Econometrica 52(2), pp. 345–362. Evans, D., 2003. The Antitrust Economics of Multi-Sided Markets. Yale Journal on Regulation 20(2), pp. 325–381. Farrell, J., 2006. Efficiency and Competition Between Payment Instruments. Review of Network Economics 5(1), pp. 26–44. Frankel, A., 1998. Monopoly and Competition in the Supply and Exchange of Money. Antitrust Law Journal 66(2), pp. 313–361. Gans, J., King, S., 2003. The Neutrality of Interchange Fees in Payment Systems. Topics in Economic Analysis & Policy 3(1), pp. 1–26. Available at: www.bepress.com/bejeap/topics/vol3/iss1/art1. Page 16 of 21 Digitization of Retail Payments Garcia-Swartz, D., Hahn, R., Layne-Farrar, A., 2006. A Move Toward a Cashless Society: A Closer Look at Payment Instrument Economics. Review of Network Economics 5(2), pp. 175–198. Green, J., 2008. Exclusive Bankcard Profitability Study and Annual Report 2008. Card & Payments, pp. 36–38. Guthrie, G., Wright, J., 2007. Competing Payment Schemes. Journal of Industrial Economics 55(1), pp. 37–67. Hancock, D., Humphrey, D., Wilcox, J., 1999. Cost Reductions in Electronic Payments: The Roles of Consolidation, Economies of Scale, and Technical Change. Journal of Banking and Finance 23(2–4), pp. 391–421. Hayashi, F., Klee, E., 2003. Technology Adoption and Consumer Payments: Evidence from Survey Data. Review of Network Economics, 2(2) pp. 175–190. Hayes, R., 2007. An Economic Analysis of the Impact of the RBA's Credit Card Reforms. mimeo, University of Melbourne. Humphrey, D., 2010. Retail Payments: New Contributions, Empirical Results, and Unanswered Questions. Journal of Banking and Finance 34(8), pp. 1729–1737. Humphrey, D., Kim, M., Vale, B., 2001. Realizing the Gains from Electronic Payments: Costs, Pricing, and Payment Choice. Journal of Money, Credit and Banking 33(2), pp. 216–234. Humphrey, D., Pulley, L., Vesala, J., 2000. The Check's in the Mail: Why the United States Lags in the Adoption of Cost-saving Electronic Payments. Journal of Financial Services Research 17(1), pp. 17–39. Humphrey, D., Willesson, M., Bergendahl, G., Lindblom, T., 2006. Benefits from a Changing Payment Technology in European Banking. Journal of Banking and Finance 30(6), pp. 1631–1652. (p. 134) Jonker, N., 2007. Payment Instruments as Perceived by Consumers—Results from a Household Survey. De Economist 155(3), pp. 271–303. Kahn, C., Roberds, W., 2009. Why pay? An Introduction to Payment Economics. Journal of Financial Intermediation 18(1), pp. 1–23. Khiaonarong, T., 2003. Payment Systems Efficiency, Policy Approaches, and the Role of the Central Bank. Bank of Finland, discussion paper, No. 1/2003. Klee, E., 2008. How People Pay? Evidence from Grocery Store Data. Journal of Monetary Economics 55(3), pp. 526– 541. Kosse, A., 2010. The Safety of Cash and Debit Cards: A Study on the Perception and Behavior of Dutch Consumers. De Nederlandsche Bank, working paper, No. 245. McAndrews, J., Wang, Z., 2008. The Economics of Two-sided Payment Card Markets: Pricing, Adoption and Usage. Federal Reserve Bank of Kansas City Working Paper, RWP 08–12. Mester, L., 2006. Changes in the Use of Electronic Means of Payment: 1995–2004. Business Review, Federal Reserve Bank of Philadelphia, Second Quarter 2006, pp. 26–30. Prager, R., Manuszak, M., Kiser, E., Borzekowski, R., 2009. Interchange Fees and Payment Card Networks: Economics, Industry Developments, and Policy Issues. Federal Reserve Board Finance and Economics Discussion Series, 2009–23. Quaden, G. (presenter), 2005. Costs, Advantages, and Disadvantages of Different Payment Methods. National Bank of Belgium, report, December. Reserve Bank of Australia, 2008a. Reform of Australia's Payments System: Preliminary Conclusions of the 2007/08 Review, April. Reserve Bank of Australia, 2008b. Reform of Australia's Payments System: Conclusions of the 2007/08 Review, September. Page 17 of 21 Digitization of Retail Payments Rochet, J.-C., Tirole, J., 2002. Cooperation Among Competitors: Some Economics of Payment Card Associations. RAND Journal of Economics 33(4), pp. 549–570. Rochet, J.-C., Tirole, J., 2003. Platform Competition in Two-Sided Markets. Journal of the European Economic Association 1(4), pp. 990–1029. Rochet, J.-C., Tirole, J., 2006a. Externalities and Regulation in Card Payment Systems. Review of Network Economics 5(1), pp. 1–14. Rochet, J.-C., Tirole, J., 2006b. Two-sided Markets: A progress Report. RAND Journal of Economics 37(3), pp. 645– 667. Rochet, J.-C., Tirole, J., 2010. Tying in Two-sided Markets and Honour all Cards Rule. International Journal of Industrial Organization 26(6), pp. 1333–1347. Rochet, J.-C., Tirole, J., 2011. Must-Take Cards: Merchant Discounts and Avoided Costs. Journal of the European Economic Association, 9(3), pp. 462–495. Rochet, J.-C., Wright, J., 2010. Credit Card Interchange Fees. Journal of Banking and Finance 34(8), pp. 1788–1797. Rysman, M., 2007. An Empirical Analysis of Payment Card Usage. Journal of Industrial Economics 55(1), pp. 1–36. Schmalensee, R., 2002. Payment Systems and Interchange Fees. Journal of Industrial Economics 50(2), pp. 103– 122. Scholtes, S., 2009. Record Credit Card Losses Force Banks into Action. Financial Times, July 1. Schuh, S., Shy, O., Stavins, J., 2010. Who Gains and Who Loses from Credit Card Payments? Theory and Calibrations. Federal Reserve Bank of Boston Public Policy Discussion Papers, 10–03. (p. 135) Schuh, S., Stavins, J., 2010. Why Are (Some) Consumers (Finally) Writing Fewer Checks? The Role of Payment Characteristics. Journal of Banking and Finance 34(8), pp. 1745–1758. Schwartz, M., Vincent, D., 2006. The no Surcharge Rule and Card User Rebates: Vertical Control by a Payment Network. Review of Network Economics 5(1), pp. 72–102. Stavins, J., 2001. Effect of Consumer Characteristics on the Use of Payment Instruments. New England Economic Review, issue 3, pp. 19–31. Stix, H., 2003. How Do Debit Cards Affect Cash Demand? Survey Data Evidence. Empirica 31, pp. 93–115. Wang, Z., 2010. Market Structure and Payment Card Pricing: What Drives the Interchange? International Journal of Industrial Organization, 28(1), pp. 86–98. Weyl, E.G., 2010. A Price Theory of Multi-sided Platforms. American Economic Review, 100(4), pp. 1642–1672. Wright, J., 2004. The Determinants of Optimal Interchange Fees in Payment Systems. Journal of Industrial Economics 52(1), pp. 1–26. Zinman, J., 2009. Debit or Credit? Journal of Banking and Finance, 33(2), pp. 358–366. Notes: (1.) For a recent discussion about the role of public authorities in retail payment systems, see Bolt et al. (2010). (2.) There are two types of payment card networks—open (four-party) and proprietary (three-party) networks. Open networks allow many banks to provide payment services to consumers and merchants, whereas in proprietary networks, one institution provides services to both consumers and merchants. When the issuer is not also the acquirer, the issuer receives an interchange fee from the acquirer. Open networks have interchange fees, whereas proprietary systems do not have explicit interchange fees because one institution serves both consumers Page 18 of 21 Digitization of Retail Payments and merchants using that network's payment services. However, proprietary networks still set prices for each side of the market to ensure that both sides are on board. (3.) Although no surcharge rules are present in the United States, merchants are free to offer cash discounts along with discounts for other payment instruments. A section of Title 10 of the Dodd-Frank Wall Street Reform and Consumer Protection Act expands the ability of merchants to provide discounts for types of payment instruments such as debit cards, checks, cash, and other forms of payments. (4.) In some instances, merchants are charged a fixed fee and a proportional fee. (5.) There are countries, for example, France, where the cardholder's account is debited much later. These types of cards are referred to as delayed debit cards. (6.) There are transaction accounts that allow overdrafts. Such accounts are common in Germany and the United States. While these types of transactions are similar to credit card transactions, however without interest-free “grace” periods when monthly balances are paid in full, potential cross-subsidies between merchants and consumers are generally more limited. (7.) For a summary of the US prepaid card market, see Chakravorti and Lubasi (2006). (8.) For a brief survey on the empirical payment literature, see Kahn and Roberds (2009). (9.) Some key benefits of using cash include privacy and anonymity that are not provided by other types of payment instruments. (10.) Carbó-Valverde et al. (2008) conduct a similar exercise for Spain, and find that when summing net costs and benefits across participants, debit cards are the least costly and checks are the most costly, with credit cards and cash ranking second and third, respectively. (11.) For empirical estimation of scale and scope economies resulting from Fedwire consolidation, see Hancock, Humphrey and Wilcox (1999) and Adams, Bauer and Sickles (2004). (12.) SEPA applies to all countries where the euro is used as the common currency. The implementation of SEPA started in January 2008 with the launching of the SEPA credit transfer scheme and should be completed when all national payment instruments are phased out; these instruments may not be entirely phased out until 2013. For a first theoretic analysis of SEPA and payment competition, see Bolt and Schmiedel (2011). (13.) His results were somewhat biased because the cost of labor across countries was not specified in the analysis. (14.) For a more general treatment of two-sided markets, see Armstrong (2006), Rochet and Tirole (2006b), and Weyl (2010). (15.) For a review of the academic literature on two-sided payment networks, see Bolt and Chakravorti (2008b). (16.) For more details, see Bolt and Tieman (2008). (17.) In this section, we build upon Chakravorti (2010) and Rochet and Tirole (2006a). (18.) He considers an environment where consumers are homogeneous, merchants are perfectly competitive, and the market for issuing and acquiring payment cards are competitive. (19.) For more discussion about no-surcharge rules and discounts, see Chakravorti and Shah (2003). (20.) Frankel (1998) refers to merchants’ reluctance to set different prices even when they are allowed to do so as price cohesion. (21.) More recently, since cash may be more expensive from a social viewpoint, McAndrews and Wang (2008) argue that card users may ultimately be subsidizing cash users as opposed to the standard view that cash users subsidize card users (Carlton and Frankel, 1995, Schuh, Shy, and Stavins, 2010, and Schwartz and Vincent, 2006). Page 19 of 21 Digitization of Retail Payments (22.) Schwartz and Vincent (2006) relax the common assumption made in the literature that the demand for the consumption good is fixed. However, they assume that consumers are exogenously divided into cash and card users and cannot switch between the groups. (23.) In a theoretic model, Bedre-Defolie and Calvano (2010) stress that consumers make two distinct decisions (membership and usage) whereas merchants make only one (membership). (24.) For more discussion, see Evans (2003). (25.) Donze and Dubec (2009) discuss the impact of collective setting of interchange fees and downstream competition in the ATM market. (26.) We limit our focus here to consumption credit. Payment credit—the credit that is extended by the receiver of payment or by a third party until it is converted into good funds—is ignored. For more discussion, see Chakravorti (2007). (27.) The empirical literature on credit cards has suggested interest rate stickiness along with above-market interest rates, although some have argued that the rate is low compared with alternatives such as pawn shops. For more discussion, see Ausubel (1991) and Brito and Hartley (1995). (28.) For a breakdown of issuer revenue percentages, see Green (2008). (29.) Farrell (2006) studies the impact of higher interchange fees on consumers who do not use cards and argues that policymakers should not ignore these redistributive effects. Wang (2010) finds that given a monopolistic network with price taking issuers and acquirers, networks tend to set higher than socially optimal interchange fees to boost card transaction value and compete with other types of payment instruments. (30.) Bourreau and Verdier (2010) and Rochet and Wright (2010) consider competition between merchant-specific private label and bank-issued general-purpose cards. (31.) McAndrews and Wang (2008) also explicitly consider card benefits and attach monetary values to them. (32.) Default rates and probability of theft will differ across countries. For Italy, Alvarez and Lippi (2009) estimate the probability of being pickpocketed at around 2 percent in 2004. For the United States, Scholtes (2009) reported that credit card default rates hit a record of more than 10 percent in June 2009. (33.) Reserve Bank of Australia (2008a), 17. (34.) For a summary of antitrust challenges in various jurisdictions, see Bradford and Hayashi (2008). (35.) For more discussion about card rewards, see Agarwal, Chakravorti, and Lunn (2010), Carbó-Valverde and Liñares Zegarra (2009), and Ching and Hayashi (2010). (36.) For more discussion, see Chang, Evans, and Garcia Swartz (2005) and Hayes (2007). (37.) On December 16, 2002, the Council of the European Union adopted Council Regulation (EC) No. 1/2003 on the implementation of the rules on competition laid down in Articles 81 and 82 of the Treaty Establishing the European Community (that is, the 1997 consolidated version of the Treaty of Rome). The new regulation came into effect on May 1, 2004. For more discussion on the EC's ruling on MasterCard, see Bolt (2008). (38.) Prager et al. (2009) review the US payment card market and consider potential regulations. (39.) For a theoretical model on honor-all-cards rules, see Rochet and Tirole (2010). (40.) For a detailed account of the merchants' position, see Constantine (2009). Wilko Bolt Wilko Bolt is an economist at the Economics and Research Division of De Nederlandsche Bank. Sujit Chakravorti Page 20 of 21 Digitization of Retail Payments Sujit "Bob" Chakravorti is the Chief Economist and Director of Quantitative Analysis at The Clearing House. Page 21 of 21 Mobile Telephony Oxford Handbooks Online Mobile Telephony Steffen Hoernig and Tommaso Valletti The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0006 Abstract and Keywords This article, which covers the recent advances in the economics of mobile telephony, first deals with specific market forces and how they shape competitive outcomes. It then investigates the regulation of certain wholesale and retail prices, closely following regulatory practice. Competition in mobile telephony is usually characterized by the presence of a fairly small number of competitors. Competition for market share is fierce, which benefits consumers through lower subscription fees. It is clear that with more networks, consumers will benefit from termination rates closer to cost. “Traffic management” technologies have made it possible to direct mobile phones to roam preferentially onto specific networks. While regulation has been employed to termination rates, and also international mobile roaming, no such intervention appears likely concerning retail prices. Mobile networks' business model may change from mainly providing calls and text messages to providing access to content to users. Keywords: mobile telephony, economics, wholesale prices, retail prices, traffic management, mobile networks, competition, regulation 1. Introduction Mobile (or cellular) communications markets have been growing at an impressive rate over the last two decades, with worldwide subscriptions increasing from several millions to five billions of users in 2010 on all continents. This growth was fueled not only by high adoption rates in developed economics, but increasingly by penetration in developing economies where often a fixed communications infrastructure was lacking. In the beginning the only purpose of mobile phones was to make and receive telephone calls to and from fixed-line and other mobile networks. In addition to being a telephone, modern mobile phones also support many additional services, such as text messages, multimedia messages, email, and Internet access, and have accessories and applications such as cameras, MP3 player, radio, and payment systems. The underlying technology has undergone regular changes. Pre-1990 networks were based on analog transmissions, after which digital standards such as GSM (global system for mobile communications) were introduced, denoted “second generation” networks. These networks were designed primarily for voice communications. As of this writing, the transition to third generation networks [e.g., universal mobile telecommunications system (UMTS) or code-division multiple access (CDMA)] is occurring, which are prepared for higher rates of data traffic. But already auctions are being run in many countries in order to hand out spectrum for fourth generation (long term evolution, LTE) services. The latter should allow transmission speeds high enough as to rival with fixed broadband connections. (p. 137) The mobile telephony market has provided fertile territory for a large number of theoretical and empirical papers in economics. Perhaps one reason is that its institutional features span so many interesting phenomena: Page 1 of 18 Mobile Telephony competition, regulation, network effects and standards (between parallel and different generations), two-sided platforms (mobile networks as platforms for media services), auctions (radio spectrum), etcetera. In this chapter we set out recent developments in the academic literature as concerns the analysis of competition and regulation in mobile telephony markets, mainly as concerns voice calls. Our central interest lies in the core principles of pricing of mobile networks and their interconnection. Issues related to other services (Internet, music, social networking), two-sided platforms, standards, and competition policy are dealt with in other chapters of this Handbook. Naturally, competition and regulation interact, so our division of this chapter in two parts needs to be justified. The first part of the chapter (“competition”) deals with specific market forces and how they shape competitive outcomes (be they good or bad for consumers or society). The second part (“regulation”) then provides an analysis of the regulation of certain wholesale and retail prices, closely following regulatory practice. Since both authors live in the European Union, policy issues relevant to Europe are at the forefront of this chapter. The economics questions we tackle are of more general interest, however, and extend beyond these boundaries. We hope that our chapter will provide the reader with the necessary tools to understand the upcoming decade of significant changes in mobile (but not only) communications markets. Needless to say, there have been earlier surveys of competition in mobile telephony and the accompanying literature, covering a wide range of different topics. We refer the reader to Laffont and Tirole (2000), Armstrong (2002), Vogelsang (2003), Gruber and Valletti (2003), and Gans et al. (2005). 2. Competition Competition in mobile telephony is usually characterized by the presence of a fairly small number of competitors (typically, two to four physical networks of different sizes). Barriers to entry are mainly due to a limited number of licenses granted by national authorities, reflecting a scarcity in the spectrum of electromagnetic frequencies that are needed to operate a mobile telephony network. This physical restriction has been overcome to some extent in recent years due to the creation of mobile virtual network operators (MVNOs), which are independent firms who do not own a physical network but rather rent airtime on existing ones. Furthermore, networks sell wholesale services (“termination”) to each other and often compete in tariffs which endogenously create network effects at the network level (rather than at the industry level). These features of the market can have (p. 138) significant short-term and long-term effects on how the market functions. In the following, we consider the setting of retail and wholesale prices, market structure, and dynamic issues. 2.1. On-Net and Off-Net Retail Pricing The year 1998 saw the publication of three seminal articles for the study of the economics of mobile communications: Armstrong (1998), Laffont et al. (1998a, LRTa), and Laffont et al. (1998b, LRTb). Of these, the former two consider nondiscriminatory or uniform pricing, that is, calls within the same network (on-net) and to other networks (off-net) are charged at the same price. With both linear and two-part tariffs, they find that call prices are set on the basis of a perceived marginal cost which is given by c = co + (1 – α)a, where co denotes the marginal cost of origination (the corresponding cost of termination is ct), α denotes the own market share of an operator, and a is the mobile termination rate (MTR) paid only to terminate those calls that are made to the rival, which has a market share of (1 – α). The latter article studies the same setup with discrimination between on-net and off-net calls, that is, “discriminatory tariffs,” where the relevant perceived marginal cost levels are con = co + ct for on-net calls and cof = co + a for off-net calls. LRTb find that, since networks’ costs are higher for off-net calls if the MTR is larger than the cost of termination (a 〉 ct), networks will charge more for off-net calls than for on-net calls. Under competition in two-part tariffs, where networks charge a fixed monthly fee plus per-minute prices for calls, networks would set call prices equal to perceived marginal cost, that is, pon = ct and pof = ct + a. Any difference between on-net and offnet prices would then be entirely explained by termination rates above cost. As LRTb have pointed out, this price differential has an impact on how networks compete. It creates “tariffmediated network effects” by making it cheaper to call recipients on the same network and thus creating incentives to be on the same network as one's primary contacts. Under the standard assumption that subscribers Page 2 of 18 Mobile Telephony of all networks call each other randomly (this is called a “uniform calling pattern”), this result implies that consumers will want to subscribe to the largest network. Therefore competition for market share will be fierce, which benefits consumers through lower subscription fees. On the other hand, it makes life harder for recent entrants, who will find it more difficult to sign up subscribers. Hoernig et al. (2010) analyze nonuniform calling patterns in this respect and find that, if calling patterns get more concentrated in the sense that subscribers mostly call “neighbors” rather than random people, this effect on competition becomes weaker because on-net (respectively off-net) prices will be distorted upward (respectively downward) due to networks’ attempt at price-discriminating against more captive consumers.1 An important recent addition to this literature is the consideration of subscribers’ utility of receiving (rather than just making) calls, which results in a positive (p. 139) externality imposed on receivers by callers. The implications for network competition were first studied in Jeon et al. (2004). They show that, as concerns on-net calls, the externality is internalized by charging the efficient price below cost. More precisely, if the utility of receiving the quantity q of calls is given by γu (q), where γε [0, 1] measures the strength of the call externality, the on-net price will be equal to pon = con/(1 + γ) 〈 con. On the other hand, off-net call prices will be distorted upward for strategic reasons: Since a receiver's utility increases with the number of calls received, each competing network tries to restrict the number of off-net calls made by increasing its off-net price. This both penalizes own customers (as they will make fewer calls) as well as rival's customers (as they will receive fewer calls). If the call externality is very strong, the latter effect prevails and the distortion in off-net call prices may become so large that a “connection breakdown” occurs, that is, no more off-net calls will be made as they are too expensive. Cambini and Valletti (2008) show that the probability of a connection breakdown is reduced if one takes into account that calls give rise to return calls, which arises when calls made and received are complements to each other. Hoernig (2007) confirms the distortion in off-net prices if networks are of asymmetric size, but also shows that larger networks charge a higher off-net price than smaller networks, under both linear and two-part tariffs. The reason is that for a large network the cost of increasing off-net prices, which consists of a corresponding compensation of own customers through lower on-net or subscription prices, is smaller, while the benefit in terms of switching consumers who want to receive calls from a large subscriber base, is larger. As a result, under a uniform calling pattern more calls will be made from the small network's customers to those of the large network than vice versa, which implies net wholesale payments from small to large networks. A further issue studied in Hoernig (2007) is whether increasing the differential between on-net and off-net prices can be used with anti-competitive intent. More precisely, he investigates whether it could be used as a means to predation, that is to reduce the competitor's profits. He finds that setting high off-net prices can indeed reduce the competitor's wholesale profits, but that it is also a rather costly means of doing so given the implied compensation to its own customers. 2.2. A Puzzle—Why Termination Rates Are Above Cost Despite What the Models Predict A large part of the literature on competition between networks is concerned primarily with their interconnection and the setting of the corresponding wholesale prices, which are denoted as call “termination rates.” Armstrong (1998), LRTa, LRTb, and Carter and Wright (1999) considered the question of whether networks could achieve collusive outcomes in the retail market by jointly choosing the termination (p. 140) rate. This research question should be seen in the light of the broader question of whether competition between firms owning communications infrastructures should involve only minimal regulation, such as an obligation to give access and negotiate over the respective charges, or whether wholesale prices should be regulated directly. A concern is that wholesale rates might be set in such a way as to relax competition in the retail market, that is, that termination rates could be used as an instrument of “tacit” collusion. What these papers found is that the answer depends on the types of tariffs used. With linear tariffs, that is, tariffs that only charge for calls made, networks would coordinate on termination rates above cost in order to raise the cost of stealing each other's clients. LRTa also consider two-part tariffs, that is, tariffs with a fixed monthly payment and additional charges for calls, with uniform pricing, and find that networks’ profits are neutral with respect to the Page 3 of 18 Mobile Telephony termination rate—any gain from changing the access charge in equilibrium is handed over to consumers through lower fixed fees. This is not to say that the termination rate is irrelevant in this case, quite on the contrary, since it continues to determine the level of retail prices. The puzzle arises when one considers the final and practically very relevant case, that of two-part tariffs under termination-based discrimination between on-net and off-net prices: Gans and King (2001), correcting a mistake in LRTb, show that networks would actually want to set a termination rate below cost in order to reduce the competitive intensity maintained by high on-net/off-net differentials and the resulting tariff-mediated network effects. Furthermore, Berger (2004) showed that if call externalities are strong, then networks will want to choose a termination rate below cost even if competition is in linear discriminatory tariffs. In contrast with these latter findings, market participants and sectoral regulators in particular have repeatedly voiced concern that unregulated wholesale charges for mobile termination are too high (and significantly above cost). A simple explanation for this fact may be that networks may set their termination rates unilaterally, that is, by maximizing profits over their own termination rate without actually talking to their competitors, as discussed for example in Carter and Wright (1999). This corresponds to a typical situation of “double marginalization,” where two simultaneous margins at the wholesale and at the retail level will lead to a very inefficient outcome. The “market analysis” performed under the European Regulatory Framework for Communications adheres to this logic: Each (mobile) network is a monopolist on termination of calls to its own customers and therefore has the market power to raise wholesale prices significantly above cost. Even so, there have been a number of approaches that show that mobile networks may want to set termination rates above cost even if they negotiate with each other as might have been imposed as a “light” regulatory obligation. This is, for instance, the approach followed in the United States, where under the 1996 Telecommunications Act operators must negotiate pair-wise reciprocal termination (p. 141) rates. In all the cases outlined below, it is the interaction with additional economic effects that makes networks choose a high termination rate, while the underlying model is the same as in Gans and King (2001). We now turn the discussion to these additional effects, which include termination from both fixed and mobile calls (section 1.2.1), differences among customers in the number of calls they make or in the number of calls destined to them (section 1.2.2), and impact of termination rates on market structure (section 1.2.3). 2.2.1. Arbitrage One issue that had been neglected, on purpose, in the work on two-way interconnection was interconnection with the fixed network. The economics of the latter is in fact quite different, as pointed out by Gans and King (2000), Mason and Valletti (2001), and Wright (2002). The fixed network, most of the times, is heavily regulated in the sense that it is forced to interconnect with mobile networks and must charge a termination rate for incoming calls at cost-oriented rates an order of magnitude below current mobile termination rates. This implies that mobile networks have all the bargaining power in setting their termination rates for fixed-to-mobile calls.2 Competition between mobile networks for subscribers does not at all lower fixed-to-mobile termination rates. Rather, networks attempt to maximize the corresponding profits received from calls initiated and paid by fixed users calling their mobile subscribers. Hence, mobile networks set high termination rates and spend part of this money on customer acquisition. If callers on the fixed network cannot identify the mobile network they are calling, then the resulting pricing externality (the network that raises its termination rate only suffers a part of the reduction in demand) leads to termination rates even above the monopoly level. A relation of two-way interconnection with this problem arises in two ways: Either both mobileto-mobile and fixedto-mobile termination rates are forced by regulation to be set at the same level, or “arbitrage” possibilities force them to be so, as discussed in Armstrong and Wright (2009). The typical case of “arbitrage” and its effects is France. Before the introduction of a unique mobile termination rate in 2005, mobileto-mobile calls were exchanged at a termination rate of zero (“bill & keep,” discussed later), while fixed-to-mobile termination rates where high. The discrepancy in the rates attracted arbitrageurs, using the socalled “GSM gateways.” Fixed operators could cut their costs by routing all the fixed-to-mobile traffic via a GSM gateway that transformed their calls into mobile-to-mobile calls and, by doing so, avoided the termination rate. Mobile networks in the end reacted to this and preferred to charge equally for both types of termination, thus preventing further arbitrage attempts. Page 4 of 18 Mobile Telephony Taking as the model for mobile-to-mobile interconnection the Gans and King (2001) framework, Armstrong and Wright (2009) analyze whether the incentives to set high or low termination rates are stronger if the rate is the same for both types of termination. They find that networks will want to set rates above cost, though somewhat lower than under pure fixed-to-mobile interconnection. (p. 142) 2.2.2. Elastic Subscription Demand and Customer Heterogeneity The basic result of Gans and King (2001) arises because of a business stealing effect: competing operators prefer termination rates below cost in order to soften competition for subscribers. This result arises in the context of “mature” markets, that is, markets with a perfectly price-inelastic demand for subscription. If, instead, total subscription demand were elastic, by setting high retail prices to consumers, firms would suffer from a reduced network externality effect. Firms may thus have a common incentive to increase market penetration, as this increases the value of subscription to each customer. Note that this second force works against the first one, since softening competition would cause a reduction in the number of subscribers. Hurkens and Jeon (2009) study this problem, generalizing earlier results in Dessein (2003), and still find that operators prefer below-cost termination rates. Allowing for heterogeneity of customers of the type analyzed by Cherdron (2002), Dessein (2003), and Hahn (2004) does not suffice to overturn this result, either. They consider situations where consumers differ in the amount of calls they place, and operators set fully nonlinear pricing schedules rather than the (simpler) two-part pricing schedules studied in the earlier literature. Heterogeneity of calling patterns, however, can produce situations where operators are better off by setting above-cost termination rates. Jullien et al. (2010) show that this can arise when there are two different groups of users, heavy users and light users, and when the light users have an elastic subscription demand. Also, light users are assumed to receive far more calls than they make. They find that, in the absence of termination-based price discrimination, firms prefer termination rates above cost because of a competition-softening effect: losing a caller to a rival network increases the profit that the original network makes from terminating calls on its light users which are receivers. Termination-based price discrimination dilutes this result, however, as it re-introduces the basic effect of Gans and King (2001). Hoernig et al. (2010) study the role of “calling clubs.” That is, people are more likely to call friends (people similar to themselves in terms of preferences for an operator) than other people. In this case, the calling pattern is not uniform as typically assumed in the literature, but skewed. They show that if the calling pattern becomes more concentrated, then firms prefer to have a termination rate above cost. Essentially, the “marginal” consumer who is indifferent between the offers of the two networks does not care much about the size of the two networks, and therefore the role of “tariff-mediated” network externalities is much watered down. A conceptually similar mechanism arises in Hurkens and Lopez (2010). They relax the assumption of rationally responsive expectations of subscribers and replace it by one of fulfilled equilibrium expectations, where the expectations of consumers about market shares do not change with price variations (off the equilibrium path). Again, this diminishes the role played by the mechanism identified by Gans and King (2001). (p. 143) Calling clubs are also studied by Gabrielsen and Vagstad (2008) and—together with switching costs— they may give a reason for operators to charge above-cost termination rates. If all members of a calling club are subscribing to the same network, price discrimination will tend to increase individual switching costs, and this may enable firms to charge higher fixed fees. To reach this result, it is essential that some customers face very high exogenous switching costs to make other people reluctant to relocate away from those friends who are locked in. In Gabrielsen and Vagstad (2008), as well as in Calzada and Valletti (2008) who also study calling clubs, subscribers with identical preferences are always part of the same calling club, and these clubs are completely separated from each other. Instead, Hoernig et al. (2010) allow for more arbitrary overlaps between calling patterns of different subscribers. Each consumer is more likely to call other consumers that are more closely located in the space of preferences, yet calls will be placed also to people that are far away. This is used to describe ties between partially overlapping social networks. 2.2.3. Foreclosure and Entry Page 5 of 18 Mobile Telephony Another reason for high termination rates that has been advanced is potentially anti-competitive practices put in place by incumbent networks. This does not only relate to a frequent complaint of recent entrants in mobile markets, as the academic literature has shown that higher termination rates can in principle be used to foreclose entrants. Essentially, in order for entry to be successful, an entrant needs termination access to the networks of its competitors and needs to be able to offer calls to these networks at reasonable prices to its own customers. High termination rates make this difficult. LRTa and LRTb argue that if entrants are small in terms of network coverage then an incumbent full-coverage network might simply refuse interconnection or charge a termination rate so high that the entrant cannot survive. Lopez and Rey (2009) revisit and broaden the analysis of the case of a monopoly incumbent facing entry, and show that the incumbent may successfully foreclose the entrant while charging monopoly prices at the retail level. However foreclosure strategies are profitable only when they result in complete entry deterrence, while using high termination rates just to limit the scale of entry, without deterring it entirely, is not profitable for the incumbent. Calzada and Valletti (2008) consider a setting similar to Gans and King (2001), with the extra twist that existing networks negotiate an industrywide (nondiscriminatory) termination rate, which will then apply also to potential entrants. Hence incumbents can decide whether to accommodate all possible entrants, only a group of them, or use the termination rate charge to completely deter entry. Incumbents thus face a trade-off. If they set an efficient (that is, industry profit maximizing) mark-up, they maximize profits ex post for a given number of firms. However, this makes entry more appealing ex ante, thus potentially attracting too many entrants and reducing profits. Faced with this threat, incumbents may want (p. 144) to distort the mark-up away from efficiency in order to limit the attractiveness of entry. 2.3. The Impact of the Size and Number of Firms While the seminal papers on network interconnection concentrated on symmetric duopoly,3 a range of additional interesting results can be obtained by allowing for asymmetrically sized networks, on the one hand, and more than two networks, on the other. Apart from being more realistic in terms of assumptions, these effects are relevant in practice but cannot be captured in symmetric duopoly models. Asymmetry can have important effects on network interconnection, but one simple fallacy should be avoided. Assuming that the large network has M subscribers and the small network N subscribers, under a balanced calling pattern the number of calls from the large to the small network is proportional to M*N, while the number of calls in the opposite direction is proportional to N*M. That is, the number of calls in either direction is equal. Thus any effects of asymmetry will arise due to additional phenomena. One issue is that asymmetry affects networks bargaining power in disputes over termination rates in different ways. With retail competition in linear tariffs the large network has high bargaining power and can impose termination rates that maintain its large market share and harm consumers, while with two-part tariffs and reciprocal termination rates the large network would prefer a termination rate equal to cost, if the asymmetry is large enough (Carter and Wright 1999, 2003, with asymmetry due to brand preferences). On the other hand, Cambini and Valletti (2004) argue that different conclusions can be reached if the asymmetry between networks arises from quality differences and affects the number of calls a customer makes. Smaller networks benefit from being able to charge higher termination charges than incumbents during the entry phase. Apart from avoiding foreclosure as mentioned above, raising the entrant's termination rate increases competitive intensity, consumer surplus, and the entrant's profits as shown in Peitz (2005a, b). On the other hand, if smaller networks’ termination rates are not regulated then they may have incentives to set the latter at excessive levels. Dewenter and Haucap (2005) show that if callers cannot distinguish which network they are calling, then due to the resulting pricing externality (the demand reduction due to a higher price is spread over all networks) a reduction in the termination rate of larger networks should lead smaller networks to further increase theirs. As concerns the setting of retail prices, Hoernig (2007) underlines that due to call externalities large networks have a strategic incentive to reduce the number of calls received by their rivals’ customers through higher off-net prices. The resulting traffic flows between networks will be unbalanced, with more call minutes emanating from smaller networks than from larger ones, which results in an “access deficit,” that is, small networks make net Page 6 of 18 Mobile Telephony access payments to larger ones. (p. 145) With respect to the number of networks, several new results obtain if one goes beyond duopoly. One question is whether the result of Gans and King that with multi-part tariffs networks would jointly agree on a termination rate below cost was robust to the presence of more networks. Calzada and Valletti (2008) show that it is, but that as the number of networks increases it will approach cost from below.4 Hoernig (2010a) considers an arbitrary number of asymmetric networks.5 Among his findings are the following: onnet/off-net differentials become smaller as the number of networks increases; while the fixed-to-mobile waterbed (see Section 3.1) is full with two-part tariffs, with linear tariffs networks retain a share of termination profits, which decreases in the number of networks and competitive intensity. An important duopoly result that is not robust to changes in the number of networks is that under multi-part tariffs consumer surplus decreases with lower termination rates: With at least three networks and sufficiently strong call externalities, consumer surplus increases when termination rates are lowered. This result is essentially due to the larger share of off-net calls. Furthermore, while it is known that in duopoly and linear tariffs higher termination rates lead to lower on-net prices, with at least three networks on-net price can increase because networks will compete less for customers. Thus with more networks it is clearer that consumers will benefit from termination rates closer to cost. 2.4. Multi-Stage Models of Interconnection There are as of yet relatively few truly dynamic studies of mobile communications markets, as opposed to the wider literature on network effects, which concentrates on competition between “incompatible” or non-interconnected networks. While most existing studies use a static model to portray a “steady state” of the market, some issues are inherently dynamic in nature. One is the dynamics of entry and market structure under network effects, and a second is the issue of incentives for network investment. Since neither can be dealt with in static models, we have grouped them in this section. 2.4.1. Market Structure Dynamics Cabral (2010) presents a dynamic network competition framework where a constant number of consumers enters and exits the market, and where depending on the circumstances market shares either equalize or the larger network will increase its dominance over time. He shows that regulated termination rates above cost increase the tendency for dominance of the larger network. Furthermore, a rate asymmetry in favor of the smaller network makes the larger network compete harder and may lower the former's discounted profits. Hoernig (2008b) considers a dynamic model of late entry into a growing communications market. He finds that tariff-mediated network effects hamper (p. 146) entry in the presence of locked-in consumers. A higher asymmetric termination rate increases the market share and discounted profits of the network concerned. Profits from fixed-to-mobile termination are fully handed over to mobile consumers in the aggregate (i.e., there is a full waterbed effect), but the large (small) network hands over less (more) than its actual termination profits due to the strategic effects of customer acquisition. Lopez (2008) considers the effect of switching costs on how networks compete for market share over time. He shows that future levels of termination rates above cost reduce earlier competition for subscribers. Thus the profit neutrality result of LRTa breaks down once one takes repeated competition into account. 2.4.2. Network Investment Cambini and Valletti (2003) and Valletti and Cambini (2005) introduce an investment stage, where firms first choose a level of quality for their networks and then compete in prices. They show that operators now have an incentive to set reciprocal termination rates above cost, in order to underinvest and avoid a costly investment battle. Abovecost termination rates make firms more reluctant to invest: the reason is that an increase of the investment in quality, relative to the rival, generates more outgoing calls and thus leads to a call imbalance which is costly precisely when termination rates are set above cost. Key to generating this result is that firms can commit to an access pricing rule prior to the investment stage. A future avenue of research that needs to be explored is how the increasing penetration of data-based retail Page 7 of 18 Mobile Telephony services impacts on competition and investment into third and fourth generation networks. Coverage could again become an issue, but the retail and interconnection pricing structures should be quite different from those in use today. While current models of competition in mobile markets capture the (largely symmetric) bi-directional nature of voice and text communications among users, the literature still lacks models that study more asymmetric data flows between, for instance, users and websites. These models should be addressed by future research aimed at discussing issues such as net neutrality, network investments, and pricing models for mobile web and wireless broadband. 3. Regulation In this section we discuss key regulatory questions that have arisen in the industry. We have organized the material as follows. First (Section 3.1), we look at the regulation of wholesale prices when mobile operators receive calls from other, noncompeting operators (most important, fixed-line users calling mobile phones). Because the termination end on the mobile networks is a bottleneck, competition problems are expected to arise there. Then (Section 3.2) we look at the problem of termination of calls between competing mobile operators. In this case, (wholesale) (p. 147) termination prices are both a source of revenues (when terminating rival traffic) and a cost (when sending traffic to the rival). In Section 3.3, we further look at the strategic impact that wholesale charges have among competing operators when they create price discrimination between on-net and off-net charges. Regulators sometime want to limit the extent of such discrimination. Finally (Section 3.4), we discuss regulatory problems related to specific access services supplied to entrants without a spectrum license, and to foreign operators to allow their customers to roam internationally. 3.1. Regulation of Fixed-to-Mobile Termination Rates: Theory and Empirics of the Waterbed Effect As anti-cipated in Section 2.2.1, in a setting where mobile operators compete against each other and receive calls from the fixed network, competition does not help to keep fixed-to-mobile (FTM) termination rates low. This situation has been called one of “competitive bottlenecks.” Mobile operators have the ability and incentives to set monopoly prices in the market for FTM calls (as the price there is paid by callers on the fixed line, not by own mobile customers), but the rents thus obtained might be exhausted via cheaper prices to mobile customers in case competition among mobile operators is vigorous. The intuition for the monopoly pricing result for the FTM market is simple: imagine FTM termination rates were set at cost; then one mobile operator, by raising its FTM termination rate, would be able to generate additional profits that it could use to lower subscription charges and attract more customers. While the mobile sector would therefore not be making any excess profits overall, an inefficiently low number of FTM calls would be made. In the absence of network externalities, Gans and King (2000), Armstrong (2002), and Wright (2002) show that the welfare maximizing level of the FTM termination rate is equal to cost. This rate level can only be achieved by direct price regulation. If, instead, there are network externalities (more specifically, benefits that accrue to fixed customers from the growth of the mobile subscriber base, as they have more subscribers that they can call), it is welfare enhancing to allow mobile operators to earn additional profits from fixed customers that can be used to attract more mobile subscribers.6 Even in this case, Armstrong (2002) and Wright (2002) show that the welfare maximizing level of the FTM termination rate would always be below the level that mobile operators would set if unregulated—thus the presence of network externalities would not obviate the need for regulation. Valletti and Houpis (2005) further extend these arguments to account for the intensity of competition in the mobile sector and for the distribution of customer preferences. Given the strong case for regulatory intervention, it is not a surprise that many countries have decided to intervene to cut these rates. Indeed, all EU member states, (p. 148) as well as several other countries, have done so, to the benefit of consumers calling mobile phones. However, reducing the level of FTM termination rates can potentially increase the level of prices for mobile subscribers, causing what is known as the “waterbed” or “seesaw” effect. The negative relationship between termination rates and prices to mobile consumers is a rather strong theoretical prediction that holds under many assumptions about the details of competition among mobile operators (Hoernig, 2010a). Over the last decade, there has been considerable variation in the toughness of regulatory intervention both Page 8 of 18 Mobile Telephony across various states and, within states, over time and among different operators (typically, entrants have been treated in a more lenient way compared to incumbents, at least in the earlier days). This has provided interesting data to test various aspects of the theories that gave intellectual support to the case for regulatory intervention. Genakos and Valletti (2011a) document empirically the existence and magnitude of the waterbed phenomenon by using a monthly panel of mobile operators’ prices across more than 20 countries over six years. Their results suggest that although regulation reduced termination rates by about 10 percent to the benefit of callers to mobile phones from fixed lines, this also led to a 5 percent increase (varying between 2 and15 percent, depending on the estimate) in mobile retail prices. They use a difference-in-difference design in the empirical specification, controlling for country-specific differences (as, say, the United Kingdom has special features different from those of Italy), for operator-specific differences (as, say, Vodafone in the UK has access to different radio frequencies than H3G UK), time-specific differences (as technological progress varies over time, as well as controlling for seasonality effects), and consumer-specific differences (as the data they use report information for different usage profiles, which may differ among themselves). After controlling for all these differences, they show that MTR regulation had a significant additional effect on retail prices set by mobile operators. They also show that the profitability of mobile firms is negatively affected by regulation, which is interpreted as evidence that the industry is oligopolistic and does not pass one-for-one termination rents to their customers. In follow-up work, Genakos and Valletti (2011b) take into account that the overall impact of regulation of termination rates will balance both effects arising from fixed-to-mobile calls and mobile-to-mobile calls. While the first effect unambiguously should push up mobile retail prices, the latter is less clear, and will depend on the type of tariff the customers subscribe to, as reviewed in section 1.2. They show that the waterbed effect is stronger for post-paid than for pre-paid contracts, where the former can be seen as an example of nonlinear tariffs, while the latter are a proxy for linear tariffs. Cunningham et al. (2010) also find evidence of the waterbed effect in a cross-section of countries. This is also the conclusion of Dewenter and Kruse (2011), although they follow an indirect approach, as they test the impact of mobile termination rates on diffusion, rather than looking directly at the impact on prices. Since the waterbed effect predicts that high MTRs should be associated with low (p. 149) mobile prices, it also predicts that diffusion will be faster in those markets with high termination rates, which is what Dewenter and Kruse (2011) find. Andersson and Hansen (2009) also provide some empirical evidence on the waterbed effect. They test the impact on the overall profitability of changes in mobile termination rates. They find that MNOs’ profits do not seem to vary statistically with changes in these MTRs. This is consistent with a hypothesis of a “full” (one-for-one) waterbed effect. The literature just reviewed falls short of predicting, from the data, what the optimal level of intervention would be, possibly because of empirical approaches taken (cross-country comparisons, rather than empirical structural models at a single-country level). This is a fruitful area for future research. An alternative is to calibrate theoretical models with realistic demand and supply parameters (see Harbord and Hoernig, 2010). 3.2. Regulation of Mobile-to-Mobile Termination Rates: Who Should Pay for Termination? The literature reviewed in Section 3.1 concentrates on the optimal setting of termination rates in one direction only. However, as interconnection between two networks typically requires two-way access, there have been some significant developments in the literature on two-way termination rates, which we reviewed in Section 2.2. The results of this research are more directly applicable to environments where operators compete for the same subscriber base (e.g., mobile-to-mobile). A relevant theme developed in this literature is “call externalities” (as opposed to “network externalities,” which were more relevant in the literature on FTM termination) resulting from the benefit a called party obtains from receiving a call, and how these should be taken into account by optimal (welfare-maximizing) regulation. DeGraba (2003) provides an analysis of off-net calls between subscribers of different networks. He shows that a “bill-and-keep” mechanism, where the reciprocal termination rate is set at zero, is a way of sharing efficiently the value created by a call when callers and receivers obtain equal benefit from it. Indeed, bill-and-keep is a system which is in place in the United States, where—in contrast to most of the rest of the world—it is also the case that mobile customers pay to receive calls (this system is called receiving party pays, RPP).7 This has led to some ambiguities in the regulatory debate addressed at finding remedies to high termination rates in Europe, where instead a calling party pays (CPP) system is in place. Sometimes, CPP and RPP have been Page 9 of 18 Mobile Telephony considered as mutually exclusive pricing systems (Littlechild, 2006), rather than tariff structures that are chosen by market participants. Cambini and Valletti (2008), following Jeon et al. (2004), show that there is a link between outgoing and incoming prices as a function of the termination rate. They argue that RPP can emerge endogenously when termination rates are set sufficiently low, for example using bill-and-keep. In other words, (p. 150) the remedy lies in the way termination rates are set, not in the pricing structure: if the termination rate is very low, then RPP might be introduced, not the other way around. Conversely, if termination rates are high, only a CPP system will emerge. Lopez (2011) confirms this result for a model where both senders and receivers decide when to hang up. Cambini and Valletti (2008) also establish the possibility of a first-best out-come when termination rates are negotiated under the “light” regulatory obligation of reciprocity. They show that under some circumstances the industry can self-regulate and achieve an efficient allocation via negotiated termination rates that internalize externalities, in particular when there is “induced” traffic (i.e., incoming and outgoing calls are complements). If instead incoming and outgoing calls are perfectly separable (as studied by Berger, 2004 and 2005; Jeon et al., 2004), it would be impossible to achieve efficiency without intervention. Another interesting attempt to see if it is possible to find less intrusive ways to regulate termination rates, and yet achieve efficient outcomes, is proposed by Jeon and Hurkens (2008). While most of the literature on two-way access pricing considers the termination rate as a fixed (per minute) price, they depart from this approach and apply a “retail benchmarking” approach to determine termination rates. They show how competition over termination rates can be introduced by linking the termination rate that an operator pays to the rivals to its own retail prices. This is a very pro-competitive rule: By being more aggressive, an operator both gains customers and also saves on termination payments. 3.3. Retail Prices: Mandatory Uniform Pricing Charging higher prices for off-net calls than for on-net calls, even if the underlying network cost is the same, creates welfare losses due to distorted off-net prices and leads to tariff-mediated network externalities which make it harder for small networks to compete. Still, these externalities can benefit consumers because they raise competitive intensity in general and result in higher consumer surplus. Thus an imposition of nondiscriminatory or uniform pricing may protect entrant networks with positive consequences in the long run but may also raise prices in the short run. In the presence of call externalities there are additional strategic effects which lead to even higher off-net prices. These effects disappear entirely under uniform pricing, thus call externalities have no strategic effect with mandatory uniform prices. There have been several attempts in practice at evaluating whether imposing such a restriction would be a useful regulatory measure. This should also be considered on the background that under the existing European Regulatory Framework for Communications (domestic) retail prices on mobile networks can-not be regulated, in particular because discriminatory pricing occurs independently of whether networks are dominant or not in the retail market. This does (p. 151) not mean, though, that such a restriction has never been tried out: When in 2006 the Portuguese Competition Authority (AdC) cleared the merger of Sonaecom and Portugal Telecom, which owned the first and third largest mobile networks, respectively, some of the undertakings accepted by AdC were meant to facilitate the entry of a new “third” mobile network. The merged entity (though evidently not its competitor Vodafone, given that this was not a case of sectoral regulation but of merger remedies) would commit to charge nondiscriminatory tariffs toward the potential new entrant during a transitory phase.8 Economic theory provides conflicting answers, though. LRTb conclude that, with linear tariffs, banning price discrimination may hurt consumers in two ways. First, with discriminatory pricing the termination rate only affects off-net prices and therefore reduces the extent of double marginalization implicit in the uniform price. Second, as pointed out above, tariff-mediated network externalities make networks compete harder and thus raise consumer surplus. Hoernig (2008a) shows that under multi-part tariffs depending on the form of demand the welfare-maximizing outcome may be achieved under any of three scenarios: (unregulated) price discrimination, uniform pricing, or an intermediate cap on the on-net/off-net differential. The reason why generically each outcome may be optimal under Page 10 of 18 Mobile Telephony some circumstances is that imposing a cap on the difference between on-net and off-net prices not only brings offnet prices down but also raises on-net prices to inefficient levels. Sauer (2010) shows that price discrimination has a detrimental effect on total welfare when network operators compete in multi-part tariffs and the total market size and termination rates are fixed. When market expansion possibilities exist, however, price discrimination becomes socially desirable. Cherdron (2002), assuming unbalanced calling patterns, shows that a ban on price discrimination makes networks choose efficient termination rates, instead of distorting them upward to reduce competitive intensity. In his model uniform pricing achieves the first-best outcome.9 Hoernig et al. (2010) show that, in a model with nonuniform calling patterns, the imposition of uniform pricing is always welfare increasing when access charges are equal to cost. Through creating tariffinduced network externalities, a difference between the price of on-net calls and that of off-net calls also affects the degree of competition: they also show that it is only when calling circles are sufficiently relevant that price discrimination can, through intensifying competition, increase consumer surplus. 3.4. Regulation of Nonessential Facilities We now turn the analysis to problems arising from the oligopolistic, and possibly collusive, nature of the mobile industry. While it is recognized that mobile operators compete to some extent against each other in retail markets,10 there are allegations that competition is more muted in wholesale markets. Two cases stand out for their practical relevance: access to rivals who do not have a spectrum license, and (p. 152) access to foreign operators when their customers roam abroad. We analyze both cases in turn. 3.4.1. Call Origination Radio spectrum is a scarce resource that limits the number of mobile network operators (MNOs) that can use their own radio spectrum to provide services. Without securing spectrum, facility-based entry cannot occur in this industry. In many countries, however, a different type of entry has been observed. In particular, MNOs have reached agreements to lease spare spectrum capacity to so-called mobile virtual network operators (MVNOs), that is, operators that provide mobile communications services without their own radio spectrum. In other countries, MVNOs have not yet emerged and only licensed MNOs compete against each other. Why do these differences arise? Can any inference be made on the intensity of competition in markets with and without MVNOs? As is well known, the monopolist owner of a bottleneck production factor who is also present in the downstream retail market may have both the incentive and the ability to restrict access to the bottleneck production factor, in order to reduce competition in the downstream retail market. An example of this would be a monopolist owner of a public switched telephone network, who may want to restrict access to its local loop through price or non-price means, in order to avoid competition on the markets of fixed telephony or broadband access. Even if he could reap the monopoly profit by setting a sufficiently high unregulated wholesale price (as the Chicago School has pointed out, but which may not be credible if wholesale prices are negotiated successively), the resulting retail prices would also be high, the very outcome that market liberalization intends to avoid. Low regulated wholesale prices will then strongly reduce the possibility of reaping monopoly rents and lead to stronger incentives to use non-price means to exclude competitors. In mobile telephony, however, there are reasons to believe that MNOs have different incentives than fixed telephony incumbents with respect to giving access to their networks. First and most important, MNOs are not monopolist providers of network access. The relevant case for mobile operators is where several incumbents, without exclusive control over an essential input, may or may not supply potential competitors. Therefore, even if a MNO denies access to its network to an entrant, there is no guarantee that the entrant will be blocked as it may obtain access elsewhere. Second, also because MNOs are not monopolists, a MNO that gives access to a MVNO will share with other MNOs the retail profit loss caused by the entrant. Third, and especially when entry cannot be blocked, it may be better for each MNO to be the one that gives access to the entrant. This allows the host MNO to earn additional wholesale revenues that at least partially compensate the loss in retail revenues (cannibalization) caused by the entrant. To address this problem in a structured way, imagine a stylized situation with two integrated oligopolists (the MNOs) that consider whether or not to grant wholesale access to an entrant (the MVNO). Each incumbent decides first and Page 11 of 18 Mobile Telephony (p. 153) independently whether or not to grant access to the MVNO, taking as given the access decision of the rival MNO. Then incumbents (and eventually the entrant if access has been granted) compete at the retail level. The solution to the access game is found by analyzing the following payoff matrix, where we imagine that incumbents are symmetric: MNO2 Access No access Access (a, a) (c, b) No access (b, c) (d, d) MNO1 Each MNO decides, independently from the other MNO, whether or not to grant access to the MVNO. The payoffs in each cell are derived according to the possible market interactions under the various outcomes. For instance, (c, b) in the top-right cell summaries the payoffs that arise when access is granted by MNO1 alone. Thus, three firms compete at the retail level, and firm MNO1 earns both at the retail and at the wholesale level (the total amounts to payoff c) while MNO2 does not provide access and its payoff b arises when competing against both the other incumbent and the entrant. Clearly, the solution to the access game depends on the entirety of payoffs under the various possible outcomes. These payoffs are affected by the characteristics of both the access providers and the access seeker and the complex competitive interactions between them. From the simple matrix above we can derive these results: • If d 〉 c, then {No access, No access} is a noncooperative Nash equilibrium of the access game. While taking an MVNO on board creates wholesale revenues, his presence also cannibalizes retail profits and drives down the equilibrium price level. If the latter effects are stronger then we have d 〉 c. • If a 〉 b, then {Access, Access} is a noncooperative Nash equilibrium of the access game. This outcome occurs when having an MVNO on board is preferable to him being located on some other network.11 This would occur principally if retail cannibalization is weak as compared to the increase in access profits. In other words, multiple equilibria arise when both a 〉 b and d 〉 c. This multiplicity is endemic in access games with “nonessential” facilities. It simply derives from the fact that there may be market situations where MNO1 reasons unilaterally along these lines: “If my rival gives access, I will too; but if my rival does not, I won’t either.” Once again, all this happens under the maintained assumption of uncoordinated competitive behavior. To establish the conditions that make a candidate noncooperative equilibrium more or less likely to happen, one needs to analyze very carefully the characteristics of both the MNOs and the MVNO. By providing access, a MNO may expand its market if the MVNO is able to appeal to previously unserved market segments. On (p. 154) the downside, the MVNO may cannibalize the MNO's sales. In addition, the entry of the MVNO may also affect the overall market interaction with the other incumbents. In general, the MVNO may make the incumbents softer or tougher competitors according to whether the MVNO affects the incumbents’ profits symmetrically or differentially (see Ordover and Shaffer, 2007; Brito and Pereira, 2010; Bourreau et al., 2011). The literature typically assumes that MNOs compete against each other. However, there is also the possibility that incumbent MNOs compete, perhaps imperfectly, at the retail level, but may co-ordinate their behavior at the wholesale level, in particular by refusing access to entrant MVNOs. This scenario has much policy relevance as it has been widely discussed by European NRAs, e.g., in Ireland, Italy, France, and Spain (in the latter case collusion at the wholesale level was indeed found, despite of no (direct) collusion at the retail level). To analyze this case, look again at the matrix above that portrays the general access game: There may be situations when {Access, Access} is a noncooperative equilibrium, but incumbents would be better off under {No access, No access}. That is, a sort of prisoner's dilemma could arise. If firms could coordinate (that is, collude at the wholesale level), they would choose this latter outcome without access. For this situation to occur, the following requirements must be jointly satisfied: Page 12 of 18 Mobile Telephony 1. Without coordinated behavior, the natural noncooperative outcome of the market would be one with access, that is, it must be that the payoff a 〉 payoff b 2. Without coordinated behavior, a noncooperative outcome without access would not arise, that is, it must be that the payoff c 〉 payoff d 3. Incumbents must be better off without access than with access, that is, the payoff d 〉 payoff a 3.4.2. International Roaming International roaming, that is, the use of mobile phones on visited networks in foreign countries, is made possible by compatible handset technologies and bilateral agreements between networks. It has been clear, though, that multiple market failures exist at both at the wholesale and at the retail level, and that retail prices for roaming calls, SMS and data usage have been set above network costs. A fundamental market failure originates from the transnational nature of roaming services. A roaming voice call is either originated or terminated on a visited foreign network, for which the customer's home operator makes a wholesale payment to the visited network. Then the customer pays a retail price to his home network. SMS and mobile data also involve the corresponding wholesale and retail payments. Then, if a national regulator attempts to protect customers in his country, he can only do this by affecting the retail prices charged by the home network— but if the corresponding wholesale prices, over which this regulator has no authority, remain high, low retail prices are not viable. Conversely, if a regulator intervenes to cut wholesale prices, this will go to the benefit of foreign users alone. (p. 155) For this reason, any feasible attempt to regulate roaming prices must involve the sectoral regulators of all countries involved, so that retail and wholesale prices can be controlled jointly. This is just what the European Union meant to achieve with the “Roaming Regulation” of 2007, which was amended in 2009.12 In its present form, it involves several transnational transparency and price control measures, both at the retail and at the wholesale level. A second fundamental problem in this market is that roaming services are sold as part of national bundles and that the large majority of mobile customers only spends a small fraction of their time traveling abroad. Therefore, customers may fail to take roaming prices into account when choosing their home operator and their bundle of calls, which implies that there is not much competition at the retail level for the provision of roaming calls. A further problem, which has started to be mitigated by technological advances, was that the choice of the visited foreign network used to be more or less random. Mobile phones were programmed to roam on the network with the strongest signal.13 While this made sense at a time when network coverage was not full, it implied that roaming phones were constantly and unpredictably switching the roaming supplier. Note that in this case customers rationally perceived roaming prices to be the same across operators. At the pricing level, this means that home networks as much as their customers could not predict which network would be used, and therefore visited networks could charge high wholesale prices without suffering a corresponding reduction in demand. “Traffic management” technologies have made it possible to direct mobile phones to roam preferentially onto specific networks, creating the opportunity for networks to select their foreign partners and enter into specific roaming agreements. Salsas and Koboldt (2004) analyze the effect of cross-border mergers and traffic management. They conclude that the latter may help create price competition at the wholesale level, but only if nondiscrimination constraints for wholesale pricing were lifted. Lupi and Manenti (2009) conclude that even then the market failure would persist if traffic could not be directed perfectly. Bühler (2010) analyzes the competitive effects of roaming alliances under various institutional settings. He develops a two-country model with two competing duopolists in each country. In a first stage, alliances are formed if joint profits of partners of an alliance exceed joint profits of these firms if the alliance is not formed. An alliance fixes the reciprocal wholesale price for roaming calls and each member of an alliance is committed to purchase the wholesale service from the partner of the alliance. Then, in a second stage, mobile network operators simultaneously set roaming prices that apply to operators who do not purchase the wholesale service from an alliance. These operators choose which supplier to buy from. Finally, in a third and last stage, operators set twopart retail tariffs. Since higher roaming prices imply higher perceived marginal costs of firms, an alliance has the means to relax competition. By forming an alliance, firms are able to commit to be less aggressive. While higher Page 13 of 18 Mobile Telephony roaming prices lead to lower retail profits and, thus, from a downstream perspective are not profitable, they lead to higher (p. 156) wholesale profits. The overall profit change is always positive for a small increase of the roaming price above marginal costs. Bühler also shows that retail prices would be even higher and social welfare lower under a nondiscrimination constraint at the wholesale level. Such a measure may mistakenly be introduced by a sectoral regulator in line with the idea that all firms should have access to wholesale products on equal terms, independent of whether they belong to an alliance. 4. Conclusions In this chapter we have considered the main competitive and regulatory issues that have occupied research about mobile communications markets for the last decade. Clearly the most important of these is the setting of termination rates. While the Gans and King “paradox” of networks wanting to set low termination rates is still partly unresolved, a consensus has emerged that termination rates closer to or even below cost are best for consumers, especially when the fixed network is also taken into account. A second issue is network effects that are created through retail pricing strategies, which have an impact on competitive intensity and market structure. While regulation has been applied to termination rates, and also international mobile roaming, no such intervention seems likely concerning retail prices. Looking forward, as mobile termination rates are brought down and convergence speeds up, with mobile offers increasingly being integrated technologically and from a pricing perspective into “quadruple-play” bundle offers, it seems possible that flat-rate (or bucket) offers will come to dominate the market, even in Europe. These tariffs may involve charging for the reception of calls, which, if the rebalancing from calling charges is done correctly, should improve efficiency. Mobile networks’ business model may change from mainly providing calls and text messages to providing access to content to users. This will not be easy, though, as networks tend to own the “pipes” but not the content. Acknowledgment We would like to thank Carlo Cambini and the Editors for very useful comments. References Ambjørnsen, T., Foros, O., Wasenden, O.-C., 2011. Customer Ignorance, Price Cap Regulation, and Rent-seeking in Mobile Roaming. Information Economics and Policy 23, pp. 27–36. Andersson, K., Hansen, B., 2009. Network Competition: Empirical Evidence on Mobile Termination Rates and Profitability. Mimeo. Armstrong, M., 1998. Network Interconnection in Telecommunications. Economic Journal 108, pp. 545–564. Armstrong, M., 2002. The Theory of Access Pricing and Interconnection. In: Cave, M., S. Majumdar, I. Vogelsang (Eds.), Handbook of Telecommunications Economics, North-Holland, Amsterdam, pp. 397–386. Armstrong, M., Wright, J., 2007. Mobile Call Termination. Mimeo. (p. 158) Armstrong, M., Wright, J., 2009. Mobile Call Termination. Economic Journal 119, pp. 270–307. Berger, U., 2004. Access Charges in the Presence of Call Externalities. Contributions to Economic Analysis & Policy 3(1), Article 21. Berger, U., 2005. Bill-and-keep vs. Cost-based Access Pricing Revisited. Economics Letters 86, pp.107–112. Binmore, K., Harbord, D., 2005. Bargaining over Fixed-to-mobile Termination Rates: Countervailing Buyer Power as a Constraint on Monopoly Power. Journal of Competition Law and Economics 1(3), pp. 449–472. Birke, D., 2009. The Economics of Networks: a Survey of the Empirical Literature. Journal of Economic Surveys Page 14 of 18 Mobile Telephony 23(4), pp. 762–793. Birke, D., Swann, G.M.P., 2006. Network Effects and the Choice of Mobile Phone Operator. Journal of Evolutionary Economics 16(1–2), pp. 65–84. Bourreau, M., Hombert, J., Pouyet, J., Schutz, N., 2011. Upstream Competition between Vertically Integrated Firms. Journal of Industrial Economics 56(4), pp, 677–713. Brito, D., Pereira, P., 2010. Access to Bottleneck Inputs under Oligopoly: A Prisoners’ Dilemma? Southern Economic Journal 76, pp. 660–677. Bühler, B., 2010. Do International Roaming Alliances Harm Consumers? mimeo. Busse, M.R., 2000. Multimarket Contact and Price Coordination in the Cellular Telephone Industry. Journal of Economics & Management Strategy 9(3), pp. 287–320. Cabral, L., 2010. Dynamic Price Competition with Network Effects. Review of Economic Studies 78, pp. 83–111. Calzada, J., Valletti, T., 2008. Network Competition and Entry Deterrence. Economic Journal 118, pp. 1223–1244. Cambini, C., Valletti, T., 2003. Network Competition with Price Discrimination: “Bill-and-Keep” Is Not So Bad after All. Economics Letters 81, pp. 205–213. Cambini, C., Valletti, T., 2004. Access Charges and Quality Choice in Competing Networks. Information Economics and Policy 16(3), pp. 411–437. Cambini, C., Valletti, T., 2008. Information Exchange and Competition in Communications Networks. Journal of Industrial Economics 56(4), pp. 707–728. Carter, M., Wright, J., 1999. Interconnection in Network Industries. Review of Industrial Organization 14(1), pp. 1–25. Carter, M., Wright, J., 2003. Asymmetric Network Interconnection. Review of Industrial Organization 22, pp. 27–46. Cherdron, M., 2002. Interconnection, Termination-based Price Discrimination, and Network Competition in a Mature Telecommunications Market. Mimeo. Corrocher, N., Zirulia, L., 2009. Me and You and Everyone We Know: An Empirical Analysis of Local Network Effects in Mobile Communications. Telecommunications Policy 33, pp. 68–79. Cunningham, B.M., Alexander, P.J., Candeub, A., 2010. Network Growth: Theory and Evidence from the Mobile Telephone Industry. Information Economics and Policy 22, pp. 91–102. DeGraba, P., 2003. Efficient Intercarrier Compensation for Competing Networks when Customers Share the Value of a Call. Journal of Economics & Management Strategy 12(2), pp. 207–230. Dessein, W., 2003. Network Competition in Nonlinear Pricing. RAND Journal of Economics 34(4), pp. 593–611. (p. 159) Dewenter, R., Haucap, J., 2005. The Effects of Regulating Mobile Termination Rates for Asymmetric Networks. European Journal of Law and Economics 20, pp. 185–197. Dewenter, R., Kruse, J., 2011. Calling Party Pays or Receiving Party Pays? The Diffusion of Mobile Telephony with Endogenous Regulation. Information Economics & Policy 23, pp. 107–117. Gabrielsen, T.S., Vagstad, S., 2008. Why is On-net Traffic Cheaper than Off-net Traffic? Access Markup as a Collusive Device. European Economic Review 52, pp. 99–115. Gans, J., King, S., 2000. Mobile Network Competition, Consumer Ignorance and Fixed-to-mobile Call Prices. Information Economics and Policy 12, pp. 301–327. Gans, J., King, S., 2001. Using “Bill and Keep” Interconnect Arrangements to Soften Network Competition. Page 15 of 18 Mobile Telephony Economics Letters 71, pp. 413–420. Gans, J., King, S., Wright, J., 2005. Wireless Communications. In: M. Cave et al. (Eds.), Handbook of Telecommunications Economics (Volume 2), North Holland, Amsterdam, pp. 243–288. Genakos, C., Valletti, T., 2011a. Testing the Waterbed Effect in Mobile Telecommunications. Journal of the European Economic Association 9(6), pp. 1114–1142. Genakos, C., Valletti, T., 2011b. Seesaw in the Air: Interconnection Regulation and the Structure of Mobile Tariffs. Information Economics and Policy 23(2), pp. 159–170. Gruber, H., Valletti, T., 2003. Mobile Telecommunications and Regulatory Frameworks. In: G. Madden (Ed.), The International Handbook of Telecommunications Economics (Volume 2), Edward Elgar, pp. 151–178. Hahn, J.-H., 2004. Network Competition and Interconnection with Heterogeneous Subscribers. International Journal of Industrial Organization 22, pp. 611–631. Harbord, D., Hoernig, S., 2010. Welfare Analysis of Regulating Mobile Termination Rates in the UK (with an Application to the Orange/T-Mobile Merger). CEPR Discussion Paper 7730. Harbord, D., Pagnozzi, M., 2009. “Network-Based Price Discrimination and ‘Bill-and-Keep’ vs. ‘Cost-Based’ Regulation of Mobile Termination Rates,” Review of Network Economics 9(1), Article 1. Hoernig, S., 2007. On-Net and Off-Net Pricing on Asymmetric Telecommunications Networks. Information Economics and Policy 19(2), pp. 171–188. Hoernig, S., 2008a. Tariff-Mediated Network Externalities: Is Regulatory Intervention Any Good? CEPR Discussion Paper 6866. Hoernig, S., 2008b. Market Penetration and Late Entry in Mobile Telephony. NET Working Paper 08–38. Hoernig, S., 2010a. Competition Between Multiple Asymmetric Networks: Theory and Applications. CEPR Discussion Paper 8060. Hoernig, S., Inderst, R., Valletti, T., 2010. Calling Circles: Network Competition with Non-Uniform Calling Patterns. CEPR Discussion Paper 8114. Hurkens, S., Jeon, D.S., 2009. Mobile Termination and Mobile Penetration. Mimeo. Hurkens, S., Lopez, A., 2010. Mobile Termination, Network Externalities, and Consumer Expectations. Mimeo. Jeon, D.-S., Hurkens, S., 2008. A Retail Benchmarking Approach to Efficient Two-Way Access Pricing: No Termination-based Price Discrimination. Rand Journal of Economics 39, pp. 822–849. (p. 160) Jeon, D.-S., Laffont, J.-J., Tirole, J., 2004. On the ‘Receiver-pays Principle’. RAND Journal of Economics 35(1), pp. 85–110. Jullien, B., Rey, P., Sand-Zantman, W., 2010. Mobile Call Termination Revisited. Mimeo. Laffont, J.-J., Rey, P., Tirole, J., 1998a. Network Competition I: Overview and Nondiscriminatory Pricing. RAND Journal of Economics 29(1), pp.1–37. Laffont, J.-J., Rey, P., Tirole, J., 1998b. Network Competition II: Price Discrimination. RAND Journal of Economics 29(1), pp. 38–56. Laffont, J.-J., Tirole, J., 2000. Competition in Telecommunications, MIT Press, Cambridge, Mass. Littlechild, S.C., 2006. Mobile Termination Rates: Calling Party Pays vs. Receiving Party Pays. Telecommunications Policy 30, pp. 242–277. Lopez, A., 2008. Using Future Access Charges to Soften Network Competition. Mimeo. Page 16 of 18 Mobile Telephony Lopez, A., 2011. Mobile Termination Rates and the Receiver-pays Regime. Information Economics and Policy 23(2), pp. 171–181. Lopez, A., Rey, P., 2009. Foreclosing Competition through Access Charges and Price Discrimination. Mimeo. Lupi, P., Manenti, F., 2009. Traffic Management in Wholesale International Roaming: Towards a More Efficient Market? Bulletin of Economic Research 61, pp. 379–407. Majer, T., 2010. Bilateral Monopoly in Telecommunications: Bargaining over Fixed-to-mobile Termination Rates. Mimeo. Marcus, J. S., 2004. Call Termination Fees: the U.S. in Global Perspective. Mimeo. Mason, R., Valletti, T., 2001. Competition in Communications Networks: Pricing and Regulation. Oxford Review of Economic Policy 17(3), pp. 389–415. Ordover, J., Shaffer, G., 2007. Wholesale Access in Multi-firm Markets: When is It Profitable to Supply a Competitor? International Journal of Industrial Organization 25, pp. 1026–1045. Parker, P.M., Röller, L.-H., 1997. Collusive Conduct in Duopolies: Multimarket Contact and Cross-Ownership in the Mobile Telephone Industry. RAND Journal of Economics 28(2), pp. 304–322. Peitz, M., 2005a. Asymmetric Access Price Regulation in Telecommunications Markets. European Economic Review 49, pp. 341–358. Peitz, M., 2005b. Asymmetric Regulation of Access and Price Discrimination in Telecommunications. Journal of Regulatory Economics 28(3), pp. 327–343. Salsas, R., Koboldt, C., 2004. Roaming Free? Roaming Network Selection and Inter-operator Tariffs. Information Economics and Policy 16, pp. 497–517. Sauer, D., 2010. Welfare Implications of On-Net/Off-Net Price Discrimination. Mimeo. Sutherland, E., 2010. The European Union Roaming Regulations. Mimeo. Valletti, T., Cambini, C., 2005. Investments and Network Competition. Rand Journal of Economics 36(2), pp. 446– 467. Valletti, T., Houpis, G., 2005. Mobile Termination: What is the “Right” Charge? Journal of Regulatory Economics 28, pp. 235–258. Vogelsang, I., 2003. Price Regulation of Access to Telecommunications Networks. Journal of Economic Literature 41, pp. 830–862. Wright, J., 2002. Access Pricing under Competition: an Application to Cellular Networks. Journal of Industrial Economics 50(3), pp. 289–315. Notes: (1.) See Birke (2009), Birke and Swann (2006), and Corrocher and Zirulia (2009) for empirical evidence on the strength of these effects. (2.) For a dissenting view, see Binmore and Harbord (2005). Indeed, the fixed network is also typically a large operator that should be able to affect the termination rate it pays to terminate calls, despite having to be subject to some regulatory oversight. This problem of bargaining over termination rates between two large operators, and in the “shadow of regulation,” still has to be studied in full. A first attempt in this direction is Majer (2010). (3.) LRTa and LRTb do study entry with marginal coverage. The work discussed in this section considers entrants with full coverage but a small number of subscribers. Page 17 of 18 Mobile Telephony (4.) See also Armstrong and Wright (2007) for a model of multiple symmetric networks with call externalities. (5.) Harbord and Hoernig (2010) apply this model to calibrate the effects of termination regulation and the 2010 merger between Orange and T-Mobile in the UK. (6.) This aspect was particularly relevant in earlier days of mobile growth, but has diminished in importance given the current high levels of mobile penetration. (7.) See Marcus (2004) and Harbord and Pagnozzi (2009). (8.) In the end the merger did not go ahead, so that this measure was never actually applied. (9.) Cambini and Valletti (2008) instead show that when mobile operators can charge for both making and receiving calls, they endogenously choose a termination rate that does not create differences in on-net and off-net prices at equilibrium when calls made and received are complements in the subscriber's utility function. (10.) Parker and Röller (1997) study the early (duopoly) days of the U.S. cellular industry, and find that prices were significantly above competitive levels. Their findings are in line with a theory of retail collusion due to crossownership and multi-market contacts. Busse (2000) shows that these firms used identical price schedules set across different markets as their strategic instruments to coordinate across markets. (11.) Note that this case is not the opposite of the previous one since the actions taken by the other player are different. (12.) The term “Regulation” here has a specific legal meaning. It is a measure adopted jointly by the European Commission and the Parliament, and has legal force in parallel to the existing Regulatory Framework. See Sutherland (2010) for a more detailed description. (13.) Ambjørnsen et al. (2011) argue that this system lead mobile operators to compete on “wrong” type of strategic variable, namely by overinvesting in particular hotspots to attract visiting traffic with the strongest signal rather than with the cheapest wholesale price. Steffen Hoernig Steffen Hoernig is Associate Professor with "Agregação" at the Nova School of Business and Economics in Lisbon. Tommaso Valletti Tommaso Valletti is Professor of Economics at the Business School at Imperial College, London. Page 18 of 18 Two-Sided B to B Platforms Oxford Handbooks Online Two-Sided B to B Platforms Bruno Jullien The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0007 Abstract and Keywords This article contributes to the understanding of business-to-business (B2B) platforms, drawing on insights from the two-sided market literature. The concept of this market refers to a situation where one or several competing platforms present services that help potential trading partners to interact. A platform that intermediates transactions between buyers and sellers is addressed, where the service consists in identifying the profitable trade opportunities. Two-sided markets provide a framework to investigate platforms' pricing strategies, and a convenient and insightful framework for exploring welfare issues. Key determinants of the competitive process are whether platforms obtain exclusivity from their clients or not, how differentiated they are, and what tariffs they can use. Multihoming may enhance efficiency, but has the potential adverse effect of softening competition. It can be concluded that while the literature has been concerned with antitrust implications, it has delivered few concrete recommendations for policy intervention. Keywords: B2B platforms, two-sided market, platforms pricing, welfare, multihoming, competition, tariffs 1. Introduction The development of digital technologies has led to drastic changes in the intermediation process, which combined structural separation of the physical and the informational dimensions of intermediation with major innovations in the latter domains. As discussed by Lucking-Reiley and Spulber (2000) e-commerce may be the source of several types of efficiency gains, including automation, better supply-chain management or disintermediation. This has led the way to the emergence of a new sector of activity online, where “info-mediators,” building on traditional Electronic Data Interchange, offer a wide range of electronic services helping buyers and sellers to find trading partners and conduct trade online.1 As of 2005, European e-Business Report (e-Business W@tch), which follows 10 sectors in the European Union, estimates that 19 percent of firms were using ICT solutions for e-procurement while 17 percent to support marketing or sales processes (in shares of employment). They also recently pointed to a surge of e-business since 2005 after some period of stabilization and cost-cutting, as well as to the key role of information technologies for innovating products. In the United States, the US Census Bureau2 estimates that ecommerce accounted in 2008 for 39 percent of manufacturers shipments and 20.6 percent of merchant wholesalers sales, while e-commerce sales are modest for retailers (3.6 percent of sales in 2008) although increasing. For manufacturers, e-commerce is widespread among sectors, ranging from 20 to 54 percent of shipments, the most active sectors being transportation equipment and beverage and tobacco products.3 A similar pattern arises for (p. 162) wholesalers although there is more disparity across sectors (from 7 to 60 percent), the most active sector being drugs (60 percent of sales), followed by motor vehicles and automotive equipment (47 percent). BtoB intermediation platforms offers a wide and diverse range of services to potential buyers and sellers. One Page 1 of 18 Two-Sided B to B Platforms category of services pertains to what can be referred to as matching services. This amounts to help members of the platform to identify opportunities to perform a profitable transaction (to find a match). The second category concerns support functions that help traders to improve on the efficiency of trade. This may range from simple billing and secured payment service up to integrated e-procurement solutions. Indeed the flexibility of electronic services has created numerous possibilities for combining these services into e-business offers. While some sites are specialized in guiding clients in finding the desired product with no intervention on the transactions4 others offer a full supply chain management service.5 In what follows I will be concerned primarily with the first dimension of the activity, namely the matching service. Moreover I will examine this activity from the particular angle of the two-sided market literature. The concept of a two-sided market refers to a situation where one or several competing platforms provide services that help potential trading partners to interact. It focuses on the fact that these activities involve particular forms of externalities between agents, namely two-sided network externalities. The platform is used by two sides and the benefits that a participant derives depend directly on who participates on the other side of the market.6 Most of the literature on two-sided markets has focused on the determination of an optimal price-structure for the services of the platforms in models where the mass of participants on the other side is a key driver of value for a participant, as will be explained below. BtoB platforms clearly have a two-sided dimension. The two sides are buyers and sellers, and a key determinant of the value of the service is the number of potential trading partners that an agent can reach. Admittedly there are other dimensions that matter as well. For instance eBay's innovation in dealing with issues of reliability and credibility with an efficient rating system has been part of its success.7 Also potential traders may care about the quality of the potential partners. The two-sided market perspective is partial but one that has proved to be extremely useful in providing new and original insights on the business strategies of platforms, and that leads the way for further inclusion of more sophisticated aspects in a consistent framework. The objective of this chapter is to present the main insights of the literature in the context of electronic intermediation and to open avenues for further developments. After an introduction to two-sided markets, I will discuss the implications for optimal tariffs in the case of a monopoly platform, including the role of up-front payments and of contingent transaction fees. Then the competitive case is discussed, with different degrees of differentiation, the distinction between single-homing and multihoming, and different business models. The third section (p. 163) is devoted to non-price issues such as tying, the design of the matching process and the ownership structure. The last section concludes. 2. An Introduction to Two-Sided Markets Apart from BtoB, examples of two-sided markets include payment card systems (Rochet and Tirole, 2001), video games (Hagiu, 2006), music or video platforms (Peitz and Waelbrock, 2006), media (Anderson and Coate, 2005) or health care (Bardey and Rochet, 2010).8 Telecommunication networks and Internet are also instances of two-sided markets that have been the object of many studies and the source of many insights.9 The literature on two-sided markets is concerned with the consequences for business strategies of the indirect network externalities that generate a well known chicken-and-egg problem: an agent on one side of the market is willing to participate to the platform activity only if he expects a sufficient participation level on the other side. Platforms’ strategies then aim at “bringing the two sides on board”10 and account for the demand externalities in the price structure. Along this line, Rochet and Tirole (2006) defines a platform to be two-sided when the profit and efficiency depends not only on the price level but also on the price structure.11 The most influential analysis of two-sided markets has been developed by Jean Tirole and Jean-Charles Rochet (2003, 2006). Starting from the analysis of payment card systems, they developed a general theory of platform mediated transactions highlighting the two-sided nature of these markets. Transactions take place on the platform when supply meets demand, so that the total volume of transaction depends on the willingness to pay of the two sides, the buyer side and the seller side. Let tb be the fee paid by a buyer per transaction and ts be the fee paid by a seller per transaction. Then the maximal number of transactions that buyers will be willing to conduct is a function Db(tb) that decreases with fee tb. One may define similarly a transaction demand by sellers Ds(ts). For a transaction to take place, a match should be found between a buyer and a seller who are willing to conduct the transaction. Page 2 of 18 Two-Sided B to B Platforms Thus the total number of transactions will increase with both Db and Ds. In the Rochet and Tirole (2003) formulation, as in most of the two-sided markets literature, the total volume of transaction is (proportional to) the product of the demands Db(tb)Ds(ts). This is the case for instance when each agent has a utility per transaction that is random, Di(ti) is the probability that a randomly selected agent is willing to trade and all trade opportunities are exhausted. This is also the case under random matching. (p. 164) For every transaction the platform receives a total fee t = tb + ts and supports a cost c. Therefore the revenue of the platform is given by: Maximizing the platform profit then yields the optimality conditions Under reasonable assumptions, this yields a unique solution. Thus optimality results from a balancing of the charges imposed on the two sides. While the total margin depends on some measure of aggregate elasticity, the contribution of each side to the profit depends negatively on its demand elasticity. For our purpose, we can reinterpret the formula as a standard monopoly Lerner index formula with a correct interpretation of the opportunity cost of increasing the fee on one side. For every agent on side i who is willing to transact, the platform is gaining not only the fee ti but also the fee tj that the other side pays. Thus every transaction of an agent on side i induces an effective cost c − tjalong with the revenue ti. Thus we can rewrite the formula as With this formulation we see that the fee on side i is the monopoly price for this side, but for an opportunity cost that accounts for the revenue generated on the other side by the participation of the agent. As we shall see, this generalizes to a more complex set-up. The price theory of two-sided markets is thus one of balancing the contributions to profit of the two sides. While the Rochet and Tirole (2003) model has been very influential in promoting the concept of two-sided market, its applicability to the issue of BtoB intermediation is limited for two reasons. The first reason is that it assumes that all potential buyers and sellers are participants to the platform and, in particular, that there is no fixed cost to participating. The second is that transactions conducted through BtoB platforms involve a transfer between the transacting parties. Whenever buyers and sellers bargain over the distribution of the surplus from transaction and are fully aware of each other transaction fees tb and ts, they will undo the price structure so that the final distribution of the surplus depends only on the total transaction fee t = tb + ts.12 For instance, in the case of a monopoly seller, the total price (price plus buyer fee) paid by the buyer depends only on the total fee t and not on the fee structure since the seller will adjust its price to any rebalancing of the fees between the two parties. According to the definition of Rochet and Tirole (2006), the market would then not be a two-sided market. (p. 165) Armstrong (2006), as well as Caillaud and Jullien (2001, 2003) developed alternative approaches based on the idea of indirect network effects and membership externalities.13 In these models, agents are considering the joint decision (possibly sequential) of participating to the platform and transacting with the members from the other side with whom they can interact on the platform. The models focus on one dimension of heterogeneity on each side, namely the costs of participating to the platform (time and effort devoted to the registration and the activity on the platform), and assume a uniform (but endogenous) benefits from transactions within each side. Some progress toward more dimensions of heterogeneity are made by Rochet and Tirole (2006) and more recently Weyl and White (2010). Page 3 of 18 Two-Sided B to B Platforms 3. A Model for Commercial Intermediation Let us consider a platform that intermediates transactions between buyers and sellers, where the service consists in identifying the profitable trade opportunities . To access the platform, the two types of agents are required to register. Once this is done they can start to look for trading partners. Suppose that prior to his participation in the identification process for profitable matches, an agent on one side considers all agents on the other side as equally likely to be a trading partner. In other words a buyer has no ex-ante preference about the identity of the sellers who participate. Suppose also that there is no rivalry between members of the same side. In particular sellers’ goods are non-rival, sellers have no capacity constraints and buyers have no financial constraints. In this set-up the willingness to participate of a buyer depends solely on the prices set by the platform and on the anticipated total value of the transactions performed. Moreover if the matching technology is neutral and the same for all, the total value of transactions is the product of the number of sellers and the expected value per seller (where the latter is itself the probability that the transaction occurs times the expected value of a successful transaction). In this context, we wish to discuss how the allocation resulting on the platform is affected by the level and the structure of the prices set by the platform. There is a variety of pricing instruments that an intermediary may rely upon for generating revenues from the transactions occurring on its platform. Describing this diversity of situations is beyond the scope of this chapter. For the purpose of discussing the two-sided nature of the intermediation activity, it is sufficient to distinguish between two types of prices. The first category includes fees that are insensitive to the volume of activity and are paid up-front by any user who whishes to use the platform services. I will refer to these as registration fees (they are sometime referred to as membership fees in the two-sided market literature). (p. 166) The registration fees will be denoted rb and rs for the buyers and the sellers respectively. Usually participation to the platform can be monitored so that registration fees can be imposed on both sides but they may be too costly to implement, in particular in BtoC or CtoC activities. The second category of pricing instruments includes those that are variable with the number and/or value of the transactions. Transaction fees are commonly used in BtoB activities and can take complex forms. For example, eBay.com charges sellers an “insertion fee” plus a “final value fee” paid only if the item is sold which depends on the closing value. Of course this requires the ability to monitor transactions which may depend on the nature of the transaction and the contractual agreements with the buyer or the seller. For instance, if the match between a buyer and a seller triggers multiple transactions, they may avoid paying fees to the platform by conducting the transaction outside the platform. In this case implementing fees proportional to the value of transaction requires a commitment by one party to report the transactions, either contractual or through long-run relationship and reputation effects. When feasible, the intermediary may thus charge fees tb and ts per transaction for respectively the buyer and the seller. As pointed above, with commercial transactions, the buyer and the seller negotiate a transfer so that we can assume that only the total fee t = tb + ts matters for determining whether a transaction occurs and how the surplus is shared. Notice that while registration fees and transaction fees are natural sources of revenue for BtoB platforms, they also rely on alternative sources of revenue. First advertising is a way to finance the activity without direct contribution of the parties. In so far that advertising is a way to induce interactions between potential buyers and advertisers, this can be analyzed with the toolkit of two-sided markets.14 Advertising will not be considered here, but it is worth pointing that electronic intermediation has also dramatically affected advertising by allowing direct monitoring of the ads impact through the click-through rate. Unlike banner advertising, it is difficult to draw a clear line between BtoB intermediation and click-through advertising. Indeed in both cases the platform adapts its response to the customer requests and behavioral pattern so as to propose some trade to him, and is rewarded in proportion to its success in generating interactions. Other revenues generating strategies subsume to bundling the service with some information goods. This is not only the case for most search engines that act as portals, but also for many support services that are offered by BtoB platforms, such as billing, accounting or any other information services for BtoB. Bundling may be profitable because it reduces transaction costs or exploits some forms of economy of scope.15 Bundling may also be part of an articulated two-sided market strategy, an aspect that we will briefly address. Notice that in some cases bundling 16 Page 4 of 18 Two-Sided B to B Platforms can be embedded into prices by defining prices net of the value of the good bundled.16 For instance if a service increases the value of any transaction by a fixed amount z at a cost cz for the platform, we can redefine the net transaction (p. 167) fee as t̃ = t − z and the net cost per transaction as c̃ = c − z + cz then the profit per transaction is t̃ − c̃ = t − c − cz. 3.1. Membership Externalities In the simplest version of the model, there is no transaction fee. For any pair of buyer and seller, let us normalize the expected value of transactions for a pair of buyer and seller to 1. We can then denote by α b the expected share of this surplus for a buyer and by α s = 1 − α b the expected profit of a seller per buyer registered. Notice that α s and α b don’t depend on the mass of sellers or buyers, which reflects the assumptions made above and which was the case in most of the two-sided market literature until recently (see however the discussion of externalities within sides below). This is in particular the assumption made by Armstrong (2006) and Gaudeul and Jullien (2001). Let us denote by Nb and Ns the number of buyers and the number of sellers on the platform. Then the expected utility from transactions that a buyer anticipates can be written as α bNs. As a result, the number of buyers registering is a function of α bNs − rb that we denote Db(rb − α bNs) where Db is non-increasing. Notice that this demand function is similar to the demand where quality is endogenous and measured in monetary equivalent. The term α bNs can thus be interpreted as a measure of quality that depends on the other side's participation level. Following this interpretation, the participation of sellers can be viewed as an input in the production of the service offered to buyers. As noticed in Jullien (2011), from the platform perspective agents have a dual nature: there are both clients buying the service and suppliers offering their contribution to the service delivered to the other side. This intuition then makes clear that since the price charged to one side will have to balance these two effects, it will tend to be lower than in one-sided markets. This is then no surprise that the price on one side may even be negative. Following a similar reasoning, the expected utility from transactions of a seller is α sNb leading to a level of registration Ds(rs − α sNb) With these demand functions, we obtain what can be seen as a canonical model for registration to the platform. The participation levels for registration fees rb and rs solve the system of equations (1) Under some reasonable assumptions the system of demand defines unique quantities Nb(rb, rs) and Ns(rb, rs) given the registration fees. Notice that the model involves a network externality between buyers: although buyers are not (p. 168) directly affected by other buyers, in equilibrium, each buyer creates a positive externality on the others through its impact on the sellers’ participation. This type of externalities is what is referred to in the concept of indirect network effects. Remark 1 When the demand is not unique, the price structure may matter for selecting a unique demand. Suppose for instance that Di(x) is zero for x non-negative but positive for x negative. Then zero participation is an equilibrium at all positive prices. However if one price is negative (say rb) and the other positive, the demands must be positive since Db(rb) 〈 0 The platform profit is then given by (for clarity we set the cost per transaction to zero from now on, c = 0) where cb and cs are the registration costs on the buyer side and the seller side respectively. Thus the two-sided platform profit and the monopoly prices are similar to those of a multi-product seller, where the products are the participations on each side. An intuitive derivation of these prices can be obtained as follows. Consider the following thought experiment: the platform reduces marginally the fee rb and raises the fee rs by an Page 5 of 18 Two-Sided B to B Platforms b s amount such that the participation of sellers remains unchanged. This is the case when the change of the fee rs is proportional to the change in buyers participation: drs/dNNs = α s. Then, since the sellers participation remains unaffected, the change in buyers participation is Following this logic, the effect on profit of this “neutralized” change in price is At optimal prices the change should have zero first-order effect implying that the optimal fee rb solves (the symmetric formula holds for the sellers side): (2) The interpretation is then that the registration fee is the monopoly price for the participation demand but for an opportunity cost cb − α sNs The opportunity cost is the cost cb net of the extra revenue that the platform can derive on the other side from the participation of the agent, which corresponds here to an increase in the registration fee by α s for each of the Ns sellers.17 Monopoly prices exhibit several interesting properties. First the presence of two-sided externalities (α b positive) tends to generate a lower price than without externality, very much like it is the case of direct network effects. The size of the effect on buyers’ registration fees depends, however, on the level of participation on the other side. The side that values the other side's participation the least should have (everything being equal otherwise) lower prices than the other side. This is because it is the most value enhancing for the platform which can leverage participation on (p. 169) the other side. Of course, this effect has to be combined with standard cost and demand elasticity considerations. A striking conclusion is that the optimal price can be negative on one side, if the cost is low. In many cases negative prices are hard to implement in which case the registration fee will be zero, possibly complemented with gifts and freebies. The theory thus provides a rational set-up to analyze the behavior of free services as the result of profit maximization. For this reason it is a natural tool in the economics of Internet and media, where free services are widespread. 3.2. Transaction Fees The canonical model abstracts from transaction fees and it is important to understand the role that such fees can play. While registration fees are neutral to transactions (apart for their effect on participation to the platform), transaction fees may have a direct effect on transactions. This is always the case if the benefits from trade are variable across transactions. Indeed, under efficient bargaining, a transaction occurs only when its total value is larger than the transaction fee. To start with, suppose that transaction fees are non distortionary so that the volume of transaction is not affected. Assuming a constant sharing rule of the transaction surplus between the seller and the buyer, transaction fee t leads to expected benefits α b(1 − t) for a buyer and α s(1 − t) for a seller. The expected revenue from transactions of the platform is then tNsNb proportional to the number of transaction. In the case of distortionary transaction fee, we can similarly write the respective expected benefits per member of the other side as α bv(t) and α sv(t) with v(0) − 1. Then v(t) is the net surplus of a pair of buyer and seller from a transaction and the probability that a pair of buyer and seller transact is v'(t).18 In the case of distortionary transactions fees, the total expected surplus for a pair of buyer and seller is v(t) + v'(t)t. A classical argument for two-part tariffs shows that when the platform can combine a transaction fee with registration fees, then the optimal transaction fee maximizes this total surplus (Rochet and Tirole, 2006; Jullien, (2007)). The general intuition is that the platform can always rebalance the prices so as to increase the surplus from transactions and maintain participation levels, implying that the welfare gains generated are captured by the platform. To see that, notice that demand functions are given by (3) Page 6 of 18 Two-Sided B to B Platforms Using we can write the profit in terms of quantities as (p. 170) With this formulation we see that the profit can be decomposed into two parts. The first part is a standard profit function for selling quantity Nb on the buyer side and Ns on the seller side. The second part is a term capturing the strategic dimension of the externality. From the profit formulation it is immediate that for any participation levels on each side, the platforms should set the fee at a level that maximizes the total expected surplus v(t) + v'(t)t.19 As mentioned above the conclusion that transaction fees should aim at maximizing the surplus generated by the activity of agents on the platform is reminiscent of similar results for two-part tariffs. From this we know that it has to be qualified when applied to a more general context. For one thing the conclusion relies on the assumption that agents are risk neutral. If a participant faces some uncertainty on future transactions, then transferring some of the platform revenue from fixed payment (registration fees) to transaction fees is a way to provide some insurance. The risk on the final utility of the agent is reduced which may raise efficiency when the agent is risk averse.20 Second, the conclusion relies on the assumption that the expected benefits from transactions are uniform across buyers or across sellers (see below). Hagiu (2004) points to the fact that the level of transaction fees matters also for the platform incentives to foster transactions. In his model there is insufficient trade so that total surplus is maximized by subsidizing trade. Running a deficit on transactions fosters more efficient trades by members of the platform, however it may hinder the platform incentives to attract new participants in a dynamic setting. Indeed once a significant level of participation is reached, any new participant raises the volume of transactions of already registered members and thus the deficit that the platform incurs on these transactions. Conversely positive transaction fees raise the gain of attracting a new member. The ability of the platform to commit on its future strategy then matters for the conclusion. A lack of commitment may lead the platform to opt for positive transaction fees. 3.3. Welfare Two-sided markets provide not only a framework to analyze platforms pricing strategies but also a convenient and insightful framework for analyzing welfare issues. Socially optimal pricing rules are derived in Jullien and Gaudeul (2001) and Armstrong (2006). These prices follow from standard arguments in welfare economics applied to networks. To discuss this, consider the case without transaction fees. Recall that the participation of an additional buyer generates an average value α s for every seller on the platform. Like for any other externality (pollution or congestion, for instance), since the buyer makes his decisions based on his own benefits solely, his participation decision is efficient when he faces a price that internalizes all the costs and benefits of other agents in the society. In the case of a two-sided market the net cost for society is the physical cost diminished by the value of the (p. 171) externality created on the other side, cb − α sNs. A similar argument on the other side shows that the allocation is efficient only when both sides face (total) prices pb = cb − α sNs and ps = cs − α bNb Thus socially efficient prices are below marginal cost, a conclusion that is in line with the conclusions for direct network effects. One may then interpret these prices as “social opportunity costs”. Notice that they coincide with the platform's opportunity costs. This is due to the fact that the expected benefit from transactions is uniform within each side. If it were not the case, the two opportunity costs would differ by a factor reflecting the traditional Spence distortion: the platform would care about the externality on the marginal member while social welfare would require to care about the externality on the average member (see the discussion of Weyl, 2010). Page 7 of 18 Two-Sided B to B Platforms This being understood, one can apply standard intuition to these opportunity costs. For instance, optimal tariffs under budget balanced conditions can be derived as Ramsey prices for the two segments that constitute the two sides, whereby the mark-up over the opportunity cost is proportional to the inverse elasticity of participation (Gaudeul and Jullien, 2001). Similar conclusions can be derived when transaction fees can be used (see Jullien, 2007). Clearly, social welfare is maximized when the transaction fee maximizes the expected trade surplus. Therefore, subject to the caveats discussed above, there is no divergence between privately optimal (monopoly) and socially optimal transaction fees. Given the transaction fees, the analysis of socially optimal registration fees is the same as above but accounting for the fact that the externality induced by the participation of an agent on one side includes not only the value created for members on the other side but also the profit created for the platform. 3.4. Usage Heterogeneity There is little contribution on optimal tariffs when both the benefits from transactions and the costs of participation are heterogenous. A noticeable exception is Weyl (2010) which provides a remarkable analysis of tariffs with heterogeneity in the transaction values α s and α b. Then participation demand on side i = s,b for the case with no transaction fees takes a non-linear form Di(pi, Nj) that depends on the joint distribution of the participation cost and the benefit α i The intuition developed above concerning the opportunity cost of adding one more member to the platform extends. In particular, he characterizes the increase ᾶ i of the revenue per member of side j that a platform can obtain with one more member on side i keeping the participation Nj constant on side j. The opportunity cost of selling to side i is then as above ci − ᾶ iNj Weyl then shows that monopoly pricing follows standard Lerner index formulas for these opportunity costs: (p. 172) where the elasticity term is with respect to the price. He also extends the welfare analysis by showing that socially optimal registration fees are equal to a social opportunity cost derived as follows. For one more member on side i, let ᾶ i be the mean increase of the utility of members who are registered on the other side. Then the social opportunity cost can be defined as ci − á¾± jNj Weyl identifies two distortions associated with the exercise of market power in two-sided markets. One is the standard monopoly mark-up over marginal cost. The second is reflected in the fact that ᾶ j may differ from ᾶ j, which corresponds to the standard distortion in the provision of quality by a monopoly. Indeed, as already pointed, participation on one side can be viewed as a quality dimension on the other side. When deciding on the price on one side, the monopolist accounts for the effect on the marginal member on the other side and ignores inframarginal members. By contrast, a social welfare perspective requires to account for the average member. 4. Competing Platforms Competition between platforms is shaped by the chicken-and-egg nature of the activity, as the driver of success on one side is success on the other side. As pointed in Jullien (2011), each user of a platform is both a consumer of the service and an input for the service offered to the other side. Platforms’ pricing strategies then reflect the competition to sell the service, but also the competition to buy the input. As we shall see, this dual nature of competition may generate complex strategies using cross-subsidies and a departure of prices from marginal costs. Key determinants of the competitive process discussed below are whether platforms obtain exclusivity from their clients or not, how differentiated they are and what tariffs they can use. The situation where agents can register with several platforms is usually referred to as multihoming. In discussing the competitive out-come, I devote the first part to the case of exclusive dealing by buyers and sellers (single-homing), and then I discuss multihoming. Page 8 of 18 Two-Sided B to B Platforms 4.1. Overcoming the Multiplicity Issue One issue that is faced when dealing with platform competition, that is akin to competition with network effects, is the issue of multiplicity. There may be different allocations of buyers and sellers compatible with given prices, which complicates the equilibrium analysis. There are at least three ways to address this issue. Caillaud and Jullien (2003) develop a methodology to characterize the full (p. 173) set of equilibria, based on imposing specific conditions on the demand faced by a platform deviating from the equilibrium (using the notion of pessimistic beliefs discussed below). Another approach consists in focusing on situations where multiplicity is not an issue. Along this line, Armstrong (2006) assumes enough differentiation between the platforms for the demand to be well defined. Armstrong's model has proven to be very flexible and is widely used by researchers on two-sided markets.21 Ongoing work by Weyl and White (2010) relies on a particular type of non-linear tariffs introduced in Weyl (2010), named “insulating tariffs.” These tariffs are contingent on participation levels of the two sides and designed in such a way that participation on one side becomes insensitive to the other side's participation, thereby endogenously removing the source of multiplicity. The last approach is to choose among the possible demands one particular selection. For instance, Ambrus and Argenziano (2009) imposes a game theory selection criterion (coalitional rationalizability) that can be interpreted as some limited level of coordination between sides. Jullien (2011) and Hagiu (2006) introduce the concept of pessimistic beliefs that assumes that agents always coordinate on the least favorable demand for one predetermined platform. Jullien (2011) notices that, since which demand emerges depends on agents expectations, one platform will be at a disadvantage if agents view the competing platform as focal. The concept of pessimistic beliefs captures this idea in a formal way and allows to study the effect on competition and entry of the advantage derived from favorable expectations. Beyond the technical difficulty, this multiplicity issue reminds of the importance of agents expectation in the outcome of competition between platforms. There has been little work on the role of reputation in the dynamics of BtoB marketplaces, an issue that should deserve more attention (Jullien (2011) and Caillaud and Jullien (2003) can be interpreted as addressing these issues). 4.2. Competition with Exclusive Participation The main contributions on competition between platforms are Armstrong (2006) and Caillaud and Jullien (2001, 2003).22 Exclusivity refers to the fact that agents can be active only on one platform. Whether this applies to BtoB platforms depends on the concrete context. For example an industrial buyer that relies on the platform for its supply-chain management would typically have an exclusive relationship with the platform, while his suppliers may not. More generally e-procurement may require the buyer to deal exclusively with the platform. One lesson from the literature is that the nature of competition may be drastically affected by factors such as complementary services offered by the platform and whether the platforms are perceived to be differentiated or not. To illustrate this we can contrast several contributions in the context of competition with registration fees only. (p. 174) Armstrong (2006) assumes that the platforms provide enough services for the demand to be always positive (Di(0) is positive in our model) and that these services are differentiated across platform. More precisely, Armstrong's model extends the framework of Section 2 by assuming that two symmetric platforms are differentiated à la Hotelling on each side, with large enough transportation costs and full coverage of the market. With differentiated platforms, the residual demand of one platform is well defined and the intuition on opportunity costs applies. Suppose that each platform covers half of the market on each side, and one platform decides to attract one more buyer away from its competing platform. This raises its value for sellers by α s but at the same time this reduces the value of the competing platform for sellers by α s since the buyer is moving away from the competitor. Therefore the platform can raise its registration fee rs by 2α s with unchanged sellers’ demand. Given that it serves half of the market, the platform increases its profit on the seller side by 2α s × 1/2 when it attracts one more buyer. The opportunity cost of attracting a buyer is then cb − α s, and by the same reasoning it is cs − α b for a seller. Armstrong shows that the equilibrium registration fees are the Hotelling equilibrium prices for these opportunity costs: ri = ci − α j + σi where σi is the equilibrium Hotelling mark-up (the transportation cost). The conclusion is thus that prices are reduced on both sides when the markets are two-sided compared to the one-sided case (α s = α b = 0) and that the strongest effect is on the market that induces relatively more externality for the other side, a Page 9 of 18 Two-Sided B to B Platforms conclusion in line with the lessons from the monopoly case. Caillaud and Jullien (2001), by contrast, analyze the case of non-differentiated platforms that provide pure intermediation service to homogenous buyers and sellers. Suppose there is a mass 1 on each side and that when members of side j register with platform k = A, B, a member of side i obtains a utility with no other value attached to the platform's services. In this context, demands are not well defined. If the absolute value of the price differential on each side is smaller than the average value α i of transactions, then there are two possible equilibria: one where all buyers and sellers join platform A and one where they all join platform B. To fix ideas, assume that platform B faces “pessimistic beliefs” which here imposes that in the above situation all agents register with platform A. Then the most efficient competitive strategy for platform B takes the form of a divide-andconquer strategy. When platform A offers prices ri ≤ α i, a strategy for platform B consists in setting fees rs − α i − ε 〈 0 for sellers (divide) and α b + inf{0, rb} for buyers (conquer). The platform subsidizes the sellers at a level that compensates them for being alone on the platform, thereby securing their participation. This creates a bandwagon effect on the other side of the market. Hence buyers are willing to pay α b to join the platform if rb 〉 0 (since platform A has no value without sellers) or α b − rb if rb 〈 0. A divide-and-conquer strategy then subsidizes participation on one side and recoups the subsidy by charging a positive margin on the other side. The mirror strategy subsidizing buyers can be considered. Caillaud and Jullien (2001) then show that platform serves the whole market and obtains a positive profit equal to the difference between the values of transactions |α s − α b| and Caillaud (p. 175) and Jullien (2003) show that more generally when no restriction is imposed on the demand system (except some monotonicity condition), the set of equilibria consists in one platform serving the whole market and making a profit between 0 and |α s − α b|. Armstrong (2006) and Caillaud and Jullien (2003) thus lead to very different conclusions and equilibrium strategies, where the difference lies mostly in the intensity of the impact of indirect network effects on demand. In Armstrong's model, the demand is smooth and externalities raise the elasticity of the residual demand, as in Rochet and Tirole (2003) canonical model. In Caillaud and Jullien (2001, 2003) externalities generate strong band-wagon effects and non-convexities. The model of Caillaud and Jullien has also the particularity that all the value is derived from pure intermediation. Surprisingly it is this feature that allows the active platform to generate positive profit. Jullien (2011) considers the same model but assumes that there is an intrinsic value to participating on a platform even if no member of the other side is present. If this “stand-alone” value is large enough, then the analysis of divide-and-conquer strategy is the same except that the “conquer” price is α b + rb instead of α b + inf{0, rb} (the reason is that buyers would pay rb 〉 0 even if there is no seller on the platform). This simple twist has a drastic implication since it implies that with homogenous competing platforms, there doesn’t exist a pure strategy equilibrium. The reason is that for any profit rb + rs that a platform can obtain in equilibrium, its competitor can obtain rs + rb |α s − α b| with an adequate divide-and-conquer strategies. What to retain from this? First, as it is often the case with network effects, the existence of indirect network effects in the BtoB marketplaces tends to intensify competition, at least when intermediation is only one part of the activity of platforms. A difference with the monopoly case is that the opportunity cost of attracting a member on the platform has another component than the value created on the other side: it is the “competitive hedge” due to the fact that attracting a member from the other platform deprives the latter from the very same value. Thus in a competitive context the effect of network externalities is doubled, which is what explains the inexistence result of Jullien (2011). Second, where the business models are fundamentally based on pure intermediation activities, it may be very difficult to compete with well established platforms that benefit from reputation effects. This conclusion for instance shades some light on the high degree of concentration that characterizes the market for search engines (however this should be pondered by multihoming considerations discussed below).23 New entrants may then rely on alternative strategies involving horizontal differentiation and multihoming by some agents, as discussed below. Caillaud and Jullien (2001, 2003) also illustrate the fact that the pattern of prices in two-sided markets may exhibit implicit cross-subsidies. As a result, even two sides that have similar characteristics may face very different prices, a feature that is referred to as price skewness by Rochet and Tirole (2003).24 Unlike the case of a one-sided Page 10 of 18 Two-Sided B to B Platforms market where competition tends to align prices with marginal costs, competition between two-sided platforms tends to exacerbate the skewness (p. 176) of prices and leads to permanent departure of prices from marginal costs (or even from opportunity costs as defined before). The reason is that it raises the incentives to gain a competitive advantage by courting some side. Finally, it is worth mentioning that there is much to learn for the case where the value of transactions is heterogenous within each side.25 Ambrus and Argenziano (2009) provides the insight that when agents differ in their usage value, then size on one side may act as a screening device on the other side. This opens the possibility of endogenous differentiation whereby multiple platforms with very different participation patterns coexist, the equilibrium allocation being supported by endogenous screening. 4.3. Multihoming The previous discussion assumes that agents register only to one platform which appears unrealistic for many BtoB platforms. For instance, in e-tendering, typically buyers single-home but sellers may be active on several platforms simultaneously. More generally, platforms offering supply-chain management may have only one side singlehoming depending on which end of the supply-chain is targeted by the platform's process. When one side can register to both platforms and not the other one, then a particular form of competition emerges that is referred to as “competitive bottleneck” by Armstrong (2006), where the term is borrowed from the debate over termination charges in mobile telecommunications. Indeed platforms do not really compete to attract multihoming agents as the “products” offered to them are non-rival. To see that suppose sellers are multihoming and not buyers. Then a platform can be viewed as providing an exclusive access to its registered buyers: from the perspective of sellers it is a bottleneck. Thus it can charge sellers the full value of accessing its population of registered buyers. There is no rivalry on the sellers’ side as access to one platform is not substitutable with access to another platform.26 The profit on the seller side is however competed away on single-homers. Indeed following the logic developed in the previous sections, the opportunity cost of attracting a buyer discounts the profit that the buyer allows to generate on the seller side. The equilibrium prices on the buyer side will thus be lowered to an extent that depends on the rate at which costs are passed on to buyers in equilibrium. This effect, according to which higher revenue per user on one side translates into lower prices on the other side, is what is referred to as the waterbed effect in the telecommunication literature and sometime as the seesaw effect in the literature on twosided markets.27 Notice that in comparison to single-homing, multihoming leads to higher prices for the multihoming side, there is no clear comparison on the single-homing side. The reason is that while the revenue that is generated by an additional single-homing agent is higher, the additional “competitive hedge” effect discussed in the previous section disappears with vanishing competition on the multihoming side. (p. 177) Both Armstrong (2006) and Caillaud and Jullien (2003) however concludes that profits are higher when there is multihoming on one side. Armstrong and Wright (2007) shows that a competitive bottleneck may emerge endogenously when differentiation is low on one side, but its sustainability is undermined by the possibility to offer exclusive contracts to multihomers. 4.4. Transaction Fees The results discussed so far concern competition with registration fees, which apply to some marketplaces but not all. Caillaud and Jullien (2003) analyze the outcome of Bertrand type competition with transaction fees in a context where transaction fees are non-distortionary. Much remains to be done for situation where transaction fees have distortionary effects. One main insight follows from the remark that transaction fees act as a form of risk sharing between the platform and the agents, because the payment is made only in case of a transaction while the platform would support the full cost in case no transaction occurs. Therefore they are natural tools for competing in situations involving the chicken-and-egg problem that characterizes two-sided markets. It is easier to attract a member facing uncertainty on the other side's participation if this member's payment is contingent on the other side's participation.28 In the context of Bertrand competition with single-homing, platforms would charge maximal transaction fees and subsidize participation. The competitive effect is then sufficient for the equilibrium to be efficient with zero profit.29 Page 11 of 18 Two-Sided B to B Platforms In the context of multihoming, the conclusions are more complex. Typically multihoming modifies the analysis of divide-and-conquer strategies. Indeed it is easier to convince an agent on one side to join since it does not need to de-register from the other platform. But there is a weaker band-wagon effect and it becomes more difficult to convince the other side to de-register from the competitor. In this context transaction fees play two roles. First, they raise the subsidy that can be paid upfront in order to attract an agent by deferring the revenue to the transaction stage. Second, they are competitive tools since agents may try to shop for the lowest transactions fees. As a consequence two alternative strategies can be used by platforms to conquer the market, depending on the prices of the competing platform. One strategy aims at gaining exclusivity through low registration fees or subsidies and generating revenue with high transaction fees. The other strategy consists in inducing multihoming but attracting transactions with low transaction fees. To summarize, the fact that multihoming agents try to concentrate their activity on the low transaction fee platforms creates two levels of competition. Intermediaries compete to attract registrations, and in a second stage they compete to attract transactions by multihomers. This competition tends to reduce transaction fees. One should thus expect platforms to charge lower transaction fees if there is a large extent of multihoming. In the context of Bertrand competition analyzed by Caillaud an Jullien (2003) the consequence is that an efficient equilibrium exists (p. 178) and involves zero profit if the intermediation technologies are identical.30 With different and imperfect technologies however, profits will be positive unlike the single-homing case.31 Moreover, this efficient equilibrium may coexist with inefficient competitive bottleneck equilibria. 5. Design and Other Issues While the two-sided market literature has to a large extent focused on pricing issues with indirect network effects, there are clearly other important dimensions for BtoB platforms. The literature is still in its infant stage, but some contributions address some other issues. 5.1. Sellers’ Rivalry Concerning the pricing strategies of a platform, the first obvious issue is that most of the literature abstracts from externalities between members of the same side. In the context of BtoB marketplaces the assumption is questionable as sellers of substitutable products will compete on the market, which generates negative pecuniary externalities. An analysis of two-sided markets with negative network effects within sides and positive externalities between sides is provided by Belleflamme and Toulemonde (2009) where they show that when they are neither too large or too small, negative externalities within sides impede the ability of divide-and-conquer strategies to overcome the chicken-and-egg problem. This suggests that it may be more difficult for a potential entrant to find his way to successful entry. Baye and Morgan (2001) focus more explicitly on BtoB platforms and show that when sellers are offering substitutable goods and are the source of revenue of the platform, the platform will restrict entry of sellers so as to reduce competition on the platform and preserve a positive margin for sellers (see also White, 2010).32 Hagiu (2009) shows that, as a result of reduced sellers’ competition, an increase in consumers’ preference for variety raises the relative contribution of sellers to a monopoly platform revenue.33 The paper by Nocke, Peitz and Stahl (2007) discussed below also allows for sellers’ rivalry. 5.2. Tying Traditional business strategies need to be reconsidered in the context of two-sided markets. For instance, tying has raised some attention, in part as a consequence of the debates over the recent antitrust procedures surrounding Microsoft's tying (p. 179) strategies.34 First, the traditional analysis of tying as an exclusionary practice (Whinston, 1990) needs to be reconsidered as the underlying strategic effects are complex (Hagiu and Farhi, 2008, Weyl, 2008). Second, tying may have beneficial effects specific to two-sided markets.35 For instance, Choi (2010) points to the fact that with multihoming, tying may raise the level of participation by inducing more agents to multihome, raising the global volume of transactions in the market.36 Amelio and Jullien (2007) view tying as a substitute for monetary subsidies when the latter are not feasible, thereby helping the platform to coordinate the two sides and to be more efficient in a competitive set-up, and analyze the strategic implications.37 5.3. Designing Intermediation Services Page 12 of 18 Two-Sided B to B Platforms Tying is one instance of strategic decisions by platforms that relates to the design of the platform architecture as well as pricing strategies. The development of BtoB marketplaces has led to a burgeoning of innovation in design and clearly design is an integral part of the business strategies adopted by the platform. Despite a large literature on matching design, there is little contribution on the interaction between design and pricing.38 Still this seems to be an important avenue for future research as the incentives of a platform to improve the efficiency of the intermediation process will depend on its ability to generate revenue from the improvement. Hence pricing models interact with technical choices in a non-trivial way. For instance Hagiu and Jullien (2011) provides several rationales for an intermediary offering directed matching services to direct buyers toward sub-optimal matches, and analyze the impact of pricing models on this phenomenon. Eliaz and Spiegler (2010) and de Corniere (2010) identify a mechanism in a sequential search model whereby a degradation of the quality of a pool of listed sellers leads to higher prices charged by sellers, more clicks and more revenue for the platform. For search engines, White (2009) analyzes the optimal mix of paying and non-paying listed sellers, when there are pecuniary externalities between sellers. All these contributions provide micro-foundations for design choices by platforms that affect the perceived quality of the intermediation service by each side in different ways. At a more theoretical level, Bakos and Katasamas (2008) point to the effect of vertical integration on the incentives of a platform to bias the design in favour of one side or the other. 5.4. Vertical Integration and Ownership Vertical integration is a common feature of BtoB platforms and one that is important for two reasons.39 First, vertical integration is one way to reach a critical size on one side, thereby gaining enough credibility to convince other agents to join the (p. 180) platform. Second, vertical integration leads to internalization of the surplus of the integrated buyers or sellers. It thus affects the pricing strategy of the platform and may lead them to be more aggressive in attracting new members.40 Moreover entry by integrated platforms may be more difficult to deter than entry by independent platforms, as shown for instance by Sülzle (2009). Another issue relates to the distribution of ownership which is discussed by Nocke, Peitz and Stahl (2007), where it is shown that for strong network effects, an independent concentrated ownership dominates in terms of social welfare dispersed ownership as well as vertical integration with a small club of sellers. 5.5. Sellers’ Investment While most contributions assume that the products are exogenous, some discuss the link between the platform's strategy and that sellers’ quality choices. Belleflamme and Peitz (2010) analyze sellers’ pre-affiliation investment with two competing platforms. They compare open access platforms with for-profit pay platforms and conclude that investment incentives are higher in the latter case whenever sellers’ investment raises consumers surplus to a sufficient extent (the precise meaning of “sufficient” depends on the singlehoming/multihoming nature of participation on both sides). Hagiu (2009) shows that charging transaction fees may help a platform fostering sellers’ pre-affiliation investment incentives (the argument follows from Hagiu (2004) discussed in Section 3.2). Hagiu (2007) argues that a two-sided platform may outperform traditional buy-and-resell intermediation when sellers must invest to raise consumers utility after the affiliation/wholesale decisions are made. 6. Conclusion Online intermediaries can be seen as platforms where trading partners meet and interact. The literature on twosided markets provides a useful perspective on the business strategies by focusing on two-sided network externalities and their implications for tariffs and competition. Beyond the overall price level, it is the entire price structure that matters and the literature helps understanding how the prices are affected by indirect networks effects. A key lesson is that prices should and will involve some form of cross-subsidy. Typically, the platform should court more the low externality side than the other. Moreover, unlike one-sided activities, competition exacerbates the tendency to cross-subsidy. Multihoming may improve efficiency, but has the potential adverse effect of softening competition. Much remains to be understood about competition, in particular due to the lack of a tractable well articulated model of dynamic competition. In particular the (p. 181) literature so far does not provide a clear view on barriers to Page 13 of 18 Two-Sided B to B Platforms entry. While the analysis of divide-and-conquer strategies suggests that there are opportunities for new entrants, these strategies may be excessively risky and not sustainable. Moreover issues of reputation and coordination point to the existence of barriers to entry akin to those encountered in network competition. As pointed out by Jullien (2011), the intensification of competition generated by indirect network effect suggests that there are particularly strong incentives for platforms to escape competition through differentiation and excessive market segmentation, although little is known about the determinants and the nature of platform's product design. And among the various topics for future researches mentioned in the text, the most exciting and novel concerns the linkage between design and business models. To conclude, while the literature has been concerned with antitrust implications, it has delivered few concrete recommendations for policy intervention. One of the challenge for the coming years will then be to develop models helping policy makers to deal with mergers and other antitrust issues in two-sided markets. References Ambrus A. and Argenziano R. (2009): “Network Markets and Consumer Coordination,” American Economic Journal: Microeconomics 1(1):17–52. Amelio A. and Jullien B. (2007): “Tying and Freebies in Two-Sided Markets,” IDEI Working Paper, forthcoming in International Journal of Industrial Organization. Anderson S. and Coate S. (2005): “Market Provision of Broadcasting: A Welfare Analysis,” Review of Economic Studies, 72(4): 947–972. Anderson S. and Gabszewicz J. (2006): “The Media and Advertising: A Tale of Two-Sided Markets,” in V. A. Ginsburg and D. Throsby (eds.), Handbook on the Economics of Art and Culture, vol. 1: 567–614. Armstrong M. (2002): “The Theory of Access Pricing and Interconnection,” in M. Cave, S. Majumdar and I. Vogelsang (eds.), Handbook of Telecommunications Economics, North-Holland, 295–384. Armstrong M. (2006): “Competition in Two-Sided Markets,” Rand Journal of Economics 37: 668–691. Armstrong M. and Wright J. (2007): “Two-Sided Markets, Competitive Bottlenecks and Exclusive Contracts,” Economic Theory 32: 353–380. Bakos Y. and Brynjolfsson E. (1999): “Bundling Information Goods,” Management Science 45(12):1613–1630. Bakos Y. and Katsamakas E. (2008): “Design and Ownership of Two-Sided Networks: Implications for Internet Platforms,” Journal of Management Information Systems 25(2):171–202. Bardey D. and Rochet J.C. (2010): “Competition Among Health Plans: A Two-Sided Market Approach,” Journal of Economics & Management Strategy 19(2):435–451. Baye M.R. and Morgan J. (2001): “Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets,” American Economic Review 91(3):454–474. Belleflamme P. and Peitz M. (2010): “Platform Competition and Sellers Investment Incentives,” European Economic Review 54(8):1059–1076. Belleflamme P. and Toulemonde E. (2009): “Negative Intra-group Externalities in Two-sided Markets,” International Economic Review 50(1):245–272. Caillaud B. and Jullien B. (2001): ‘Competing Cybermediaries,’ European Economic Review (Papers & Proceedings) 45: 797–808. Caillaud B. and Jullien B. (2003): “Chicken & Egg: Competition among Intermediation Service Providers,” Rand Journal of Economics 34: 309–328. Page 14 of 18 Two-Sided B to B Platforms Choi J.P. (2010): “Tying in Two-Sided Markets with Multi-homing,” The Journal of Industrial Economics 58(3):607– 626. Crampes C., Haritchabalet C. and Jullien B. (2009): “Competition with Advertising Resources,” Journal of Industrial Economics 57(1):7–31. Damiano E. and Li H. (2007): “Price Discrimination and Efficient Matching,” Economic Theory 30: 243–263. (p. 184) de Cornière A. (2010): “Targeting with Consumer Search: An Economic Analysis of Keyword Advertising,” mimeo, Paris School of Economics. Farhi E. and Hagiu A. (2008): “Strategic Interactions in Two-Sided Market Oligopolies,” mimeo, Harvard Business School. Gaudeul A. and Jullien B. (2001): “E-commerce: Quelques éléments d’économie industrielle,” Revue Economique 52: 97–117 ; English version, “E-Commerce, Two-sided Markets and Info-mediation,” in E. Brousseau and N. Curien (eds.), Internet and Digital Economics, Cambridge University Press, 2007. Hagiu A. (2006): “Optimal Pricing and Commitment in Two-Sided Markets,” The RAND Journal of Economics 37(3):720–737. Hagiu A. (2007): “Merchant or Two-Sided Platforms,” Review of Network Economics 6: 115–133. Hagiu A. (2009): “Two-Sided Platforms: Product Variety and Pricing Structures,” Journal of Economics and Management Strategy 18: 1011–1043. Hagiu A. and Jullien B. (2011): “Why Do Intermediares Divert Search?” Rand Journal of Economics 42(2): 337–362. Innes R. and Sexton R. (1993): “Customer Coalitions, Monopoly Price Discrimination and Generic Entry Deterrence,” European Economic Review 37: 1569–1597. Jullien B. (2011): “Competition in Multi-Sided Markets: Divide and Conquer,” American Economic Journal: Microeconomics 3: 1–35. Jullien B. (2006): “Pricing and Other Business Strategies for e-Procurement Platforms,” in N. Dimitri, G. Spiga and G. Spagnolo (eds.), Handbook of Procurement, Cambridge University Press. Jullien B. (2007): “Two-Sided Markets and Electronic Intermediation,” in G. Illing and M. Peitz (eds.), Industrial Organization and the Digital Economy, MIT Press. Katz M. and Shapiro K. (1985): “Network Externalities, Competition and Compatibility,” American Economic Review 75: 424–440. Katz M. and Shapiro K. (1986): “Technology Adoption in the Presence of Network Externalities,” Journal of Political Economy 94: 822–841. Kaplan S. and Sawhney M. (2000): “B2B e-Commercehubs; Toward a Taxonomy of Business Models,” Harvard Business School Review 78(3):97–103. Lucking-Reyley D. and D. Spulber (2001): “Busisness-to-Business Electronic Commerce,” Journal of Economic Prespective 15: 55–68. Nocke V., Peitz M., and Stahl K. (2007): “Platform Ownership,” Journal of the European Economic Association 5(6):1130–1160. Peitz M. and P. Waelbrock (2006): “Digital Music” in: Illing, G. and M. Peitz (eds.), Industrial Organization and the Digital Economy, MIT Press. Page 15 of 18 Two-Sided B to B Platforms Rochet J.C. and Tirole J. (2003): “Platform Competition in Two-Sided Markets,” Journal of the European Economic Association 1: 990–1029. Rochet J.C. and Tirole J. (2006): “Two-Sided Markets: a Progress Report,” The RAND Journal of Economics 37: 645– 667. Rochet J.C. and Tirole J. (2008): “Tying in Two-Sided Market and the Honor-all-Cards Rule,” International Journal of Industrial Organization 26(6):1333–1347. Rysman M. (2009): “The Economics of Two-Sided Markets,” Journal of Economic Prespective 23(3):125–143. Sülzle K. (2009): “Duopolistic Competition between Independent and Collaborative Business-to-Business Marketplaces,” International Journal of Industrial Organization 27: 615–624. Weyl E.G. (2008): “The Price Theory of Two-Sided Markets,” Harvard University. (p. 185) Weyl E.G. (2010): “A Price Theory of Multi-sided Platforms.” American Economic Review 100(4): 1642–72. Weyl G. and White A. (2010): “Imperfect Platform Competition: A General Framework” mimeo, Harvard University. Whinston D. (1990): “Tying, Foreclosure, and Exclusion,” The American Economic Review 80(4):837–859. Wright J. (2003): “Optimal Card Payment System,” European Economic Review 47, 587–617. Yoo B. Choudhary V., and Mukhopadhyay T. (2007): “Electronic B2B Marketplaces with Different Ownership Structures,” Management Sciences 53(6):952–961 (p. 186) . Notes: (1.) See the survey on electronic commerce in The Economist (February 2000). (2.) 2008 E-commerce Multi-sector “E-Stats” Report, available at http://www.census.gov/econ/estats/. (3.) The value is concentrated among six sectors: Transportation Equipment, Petroleum and Coal Products, Chemical Products, Food Products, Computer and Electronic Products and Machinery Products. (4.) Nextag and Kaboodle are examples of search engines proposing a comparison of products and prices over all funishers. (5.) Examples are Sciquest, which provides procurement and supplier management processes for life sciences, and BravoSolution, a general provider of supply management. (6.) General presentations are Rochet and Tirole (2005), Jullien (2006), and Rysman (2009) (7.) See the chapter “Reputation on the Internet” by L. Cabral in this handbook. (8.) For detailed discussions, see the chapters in this handbook “Digitalization of Retail Payment” by W. Bolt and S. Chakravorti, “Home Videogame Platforms” by R. Lee, “Software Platforms” by A. Hagiu and “Advertising on the Internet” by S. Anderson. (9.) See the survey by Armstrong (2002) and the chapter “Mobile Telephony” by S. Hoernig and T. Valletti in this handbook. (10.) The alliance between Cornerbrand and Kazaa described by Peitz and Waelbroeck (2006) is a good example of such a strategy, as each side benefits from it. (11.) Since there has been notorious difficulty in providing a consensus on a formal general definition of a twosided market, I shall not try to do so here. (12.) The statement is valid for a large class of bargaining models, including models with asymmetric information Page 16 of 18 Two-Sided B to B Platforms (see Rochet and Tirole, 2006). (13.) See also Gaudeul and Jullien (2001) for a monopoly model along the same lines. (14.) See Anderson and Gabszewicz (2006), Anderson and Coate (2005), Crampes et al. (2004), and the chapter 14 by S. Anderson in this handbook. (15.) A particular form of economy of scope for electronic goods is a reduction in demand uncertainty (see Bakos and Brynjolfson, 1999). (16.) See chapter 11 by J.P. Choi in this handbook. (17.) With a cost per transaction, the opportunity costs is cb + cNs − α sNs which may be larger or smaller than cb. (18.) To see that let u be the value of a transaction and F(u) its cdf. Then which derivative is 1 − F(u) the probability that the transaction occurs. (19.) This conclusion extends to duopoly with single-homing. (20.) As an illustration, while in 2005 the leading automotive e-procurement portal, Covinsint, was relying on membership fees, its smaller competitor Partsforindustry was relying most extensively on volume related payment (Jullien, 2006). (21.) The Rochet and Tirole (2003) model also has a unique equilibrium in duopoly. (22.) The literature on competition between mobile telecommunication networks has introduced many of the relevant concepts (see Armstrong (2002) for a review). See also chapter “Mobile Telecommunications” by S. Hoernig and T. Valletti in this handbook. (23.) See chapter 9 by J. Moraga and M. Wildenbeest in this handbook. (24.) This conclusion is corrobated by Jullien's treatment of a general competitive game between multisided platforms (Jullien, 2000). (25.) See the ongoing work by Weyl and White (2010). (26.) This assumes away any resource constraint or other diminishing returns for sellers. (27.) In the context of the Hotelling model discussed by Armstrong (2006), there is a one-to-one waterbed effect so all profits from sellers are competed away on buyers. (28.) The same insight holds for Weyl's concept of insulating tariff, which may help to provide credibility to this concept for competitive situations. (29.) With distortionary transaction fees and Bertrand competition, efficiency would not be achieved but there would still be zero profits (see Jullien, 2007). (30.) By this it is meant that if one platform fails to generate a particular transaction, the other will fail as well. (31.) With an imperfect intermediation technology, a third strategy always generates positive profits: it consists in focusing on agents who failed to perform a transaction on the competing platform, by charging a high transaction fee, exploiting the last resort position. (32.) See chapter 9 by J. Moraga and M. Wildenbeest in this handbook. (33.) Results are more ambiguous for competing platforms, for reasons similar to Belleflamme and Toulemonde (2009). (34.) See chapter 11 by J.P. Choi in this handbook. (35.) See Rochet and Tirole (2004b) for a similar view on tying between two payment card systems. Page 17 of 18 Two-Sided B to B Platforms (36.) An interesting example of efficient bundling is the ability to solve the issue of buyers’ moral hazard by bundling a well-designed payment system with the matching service (see for example the BtoC website PriceMinister). (37.) Bundling may of course serve more traditional price discrimination purposes. The reader is referred to Jullien (2006) for an informal discussion of bundling in the context of e-procurement. (38.) A noticeable exception is Damiano and Li (2008). (39.) This is usually referred to as biased marketplaces, see Kaplan and Sawheny (2000). As an example Covisint is jointly owned by car manufacturers Daimler, General Motors, Ford and Renault-Nissan. (40.) See Yoo et al. (2007). Bruno Jullien Bruno Jullien is a Member of the Toulouse School of Economics and Research Director at CNRS (National Center for Scientific Research). Page 18 of 18 Online versus Offline Competition Oxford Handbooks Online Online versus Offline Competition Ethan Lieber and Chad Syverson The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0008 Abstract and Keywords This article describes the nature of competition between online and offline retailing. It first introduces some basic empirical facts that reflect the present state of online and offline competition, and then discusses the interplay of online and offline markets. Internet users are higher income, more educated, and younger than non-Internet users. Education is a sizeable determinant of who is online, even controlling for income. E-commerce enables new distribution technologies that can decrease supply chain costs, improve service, or both. The Internet has influenced the catalog of products available to consumers. The changes in demand- and supply-side fundamentals that e-commerce brings can foment substantial shifts in market outcomes from their offline-only equilibrium. These include prices, market shares, profitability, and the types of firm operating in the market. Online channels have yet to fully establish themselves in some markets and are typically growing faster than bricks-and-mortar channels. Keywords: competition, online retailing, offline retailing, e-commerce, Internet, prices, market shares, profitability 1. Introduction Amazon is arguably one of the most successful online firms. As of this writing, its market value is more than $80 billion, 60 percent higher than the combined value of two large and successful offline retailers, Target and Kohl’s, who have almost 2,900 stores between them. Jeff Bezos conceived of Amazon as a business model with many potential advantages relative to a physical operation. It held out the potential of lower inventory and distribution costs and reduced overhead. Consumers could find the books (and later other products) they were looking for more easily, and a broader variety could be offered for sale in the first place. It could accept and fulfill orders from almost any domestic location with equal ease. And most purchases made on its site would be exempt from sales tax. On the other hand, Bezos no doubt understood some limitations of online operations. Customers would have to wait for their orders to be received, processed, and shipped. Because they couldn’t physically inspect a product before ordering, Amazon would have to make its returns and redress processes transparent and reliable, and offer other ways for consumers to learn as much about the product as possible before buying. (p. 190) Amazon's entry into the bookselling market posed strategic questions for brick-and-mortar sellers like Barnes & Noble. How should they respond to this new online channel? Should they change prices, product offerings, or capacity? Start their own online operation? If so, how much would this cannibalize their offline sales? How closely would their customers see ordering from the upstart in Seattle as a substitute for visiting their stores?1 The choices made by these firms and consumers’ responses to them—actions driven by the changes in market fundamentals wrought by the diffusion of e-commerce technologies into bookselling—changed the structure of the Page 1 of 28 Online versus Offline Competition market. As we now know, Amazon is the largest single bookseller (and sells many other products as well). Barnes & Noble, while still large, has seen its market share diminish considerably. Borders is out of business. There are also many fewer bricks-and-mortar specialty bookshops in the industry and their prices are lower. In this chapter, we discuss the nature of competition between a market's online and offline segments. We take a broad view rather than focus on a specific case study, but many of the elements that drove the evolution of retail bookselling as just described are present more generally. We organize our discussion as follows. The next section lays out some basic facts about the online sales channel: its size relative to offline sales; its growth rate; the heterogeneity in online sales intensity across different sectors, industries, and firms; and the characteristics of consumers who buy online. Section 3 discusses how markets’ online channels are economically different due to e-commerce's effects on market demand and supply fundamentals. Section 4 explores how changes in these fundamentals due to the introduction of an online sales channel might be expected to change equilibrium market outcomes. Section 5 investigates various strategic implications of dual-channeled markets for firms. A short concluding section follows. 2. Some Facts Before discussing the interplay of online and offline markets, we lay out some basic empirical facts that reflect the present state of online and offline competition. 2.1. How Large Are Online Sales Relative to Offline Sales? To take the broadest possible look at the data, it is useful to start with the comprehensive e-commerce information collected by the U.S. Census Bureau.2 The Census separately tracks online- and offline-related sales activity in four major sectors: (p. 191) manufacturing, wholesale, retail, and a select set of services. The data are summarized in Table 8.1. In 2008, total e-commerce-related sales in these sectors were $3.7 trillion. Offline sales were $18.7 trillion. Therefore transactions using some sort of online channel accounted for just over 16 percent of all sales. Not surprisingly, the online channel is growing faster: nominal e-commerce sales grew by more than 120 percent between 2002 and 2008, while nominal offline sales grew by only 30 percent. As a greater fraction of the population goes online—and uses the Internet more intensively while doing so—e-commerce's share will almost surely rise. The relative contribution of online-based sales activity varies considerably across sectors, however. Looking again at 2008, e-commerce accounted for 39 percent of sales in manufacturing and 21 percent in wholesale trade, but only 3.6 percent in retail and 2.1 percent in services. If we make a simple but broadly accurate classification of deeming manufacturing and wholesale sales as business-to-business (B2B), and retail and services as business-toconsumer (B2C), online sales are considerably more salient in relative terms in B2B sales than in B2C markets. Because total B2B and B2C sales (thus classified) are roughly equal in size, the vast majority of online sales, 92 percent, are B2B related.3 That said, B2C e-commerce is growing faster: it rose by 174 percent in nominal terms between 2002 and 2008, compared to the 118 percent growth seen in B2B sectors. In terms of shares, ecommerce-related sales in B2B sectors grew by about half (from 19 to 29 percent) from (p. 192) 2002 to 2008, while more than doubling (from 1.3 to 2.7 percent) in B2C sectors over the same period.4 Page 2 of 28 Online versus Offline Competition Table 8.1 Dollar Value of Commerce by Sector and Type ($ billions) Percent gain, Manufacturing Wholesale Retail Service Total 2002 2008 2002–2008 E-commerce 751.99 2154.48 186.5 Offline 3168.65 3331.78 5.2 Fraction e-commerce 0.192 0.393 E-commerce 806.59 1262.37 56.5 Offline 3345.01 4853.79 45.1 Fraction e-commerce 0.194 0.206 E-commerce 44.93 141.89 215.8 Offline 3089.40 3817.27 23.6 Fraction e-commerce 0.014 0.036 E-commerce 59.97 146.49 144.3 Offline 4841.03 6700.97 38.4 Fraction e-commerce 0.012 0.021 E-commerce 1,663.47 3,705.23 122.7 Offline 14,444.09 18,703.81 29.5 Fraction e-commerce 0.103 0.165 Notes: This table shows the composition of sector sales by e-commerce status. Data are from the U.S. Census E-commerce Reports (available at 〈http://www.census.gov/econ/estats/〉). See text for definition of e-commerce sales. When considering the predominance of B2B e-commerce, it is helpful to keep in mind that the data classify as ecommerce activity not just transactions conducted over open markets like the Internet, but also sales mediated via proprietary networks as well. Within many B2B sectors, the use of Electronic Data Interchange as a means to conduct business was already common before the expansion of the Internet as a sales channel during the mid 1990s. While some research has looked at the use of less open networks (e.g. Mukhopadhyay, Kekre, and Kalathur, 1995), the academic literature has focused on open-network commerce much more extensively. We believe that much of the economics of the more B2C-oriented literature discussed in this paper applies equally or nearly as well to B2B settings. Still, it is useful to keep in mind the somewhat distinct focal points of the data and the literature. 2.2. Who Sells Online? Page 3 of 28 Online versus Offline Competition In addition to the variation in online sales intensity across broad sectors, there is also considerable heterogeneity within sectors. Within manufacturing, the share of online-related sales ranges from 21 percent in Leather and Allied Products to 54 percent in Transportation Equipment. In retail, less than one third of one percent of sales at Food and Beverage stores are online; on the other hand, online sales account for 47 percent of all sales in the “Electronic Shopping and Mail-Order Houses” industry (separately classified in the NAICS taxonomy as a 4-digit industry). Similar diversity holds across industries in the wholesale and service sectors. Differences in the relative size of online sales across more narrowly defined industries can arise from multiple sources. Certain personal and business services (e.g. plumbing, dentistry, copier machine repair) are inherently unsuited for online sales, though some logistical aspects of these businesses, such as advertising and billing, can be conducted online. Likewise, consumer goods that are typically consumed immediately after production or otherwise difficult to deliver with a delay (e.g., food at restaurants or gasoline) are also rarely sold online. In an attempt to explain the heterogeneity in the online channel's share of sales across manufacturing industries, we compared an industry's e-commerce sales share to a number of plausible drivers of this share. These include the dollar value per ton of weight of the industry's output (a measure of the transportability of the product; we use its logarithm), R&D expenditures as a fraction of sales (a proxy for how “high-tech” the industry is), logged total industry sales (to capture industry size), and an index of physical product differentiation within the industry. (All variables were measured at the 3-digit NAICS level.5) We did not find clear connections of industries’ e-commerce sales shares to these potential drivers in either raw pairwise correlations or in a regression framework, though our small sample (p. 193) size makes inference difficult. The tightest link was between e-commerce intensity and the logged value per ton of the industry's output. A one-standard-deviation increase in the latter was related to a roughly half-standard-deviation increase in e-commerce's sales share. The statistical significance of this connection was marginal (the p-value is 0.101), however. Forman et al. (2003) study sources of differences in online sales activity across firms by investigating commercial firms’ investments in e-commerce capabilities. They do so using the Harte Hanks Market Intelligence CI Technology database from June 1998 through December of 2000, which contains information on technology use for more than 300,000 establishments. The authors use this data to classify investments in e-commerce capabilities into two categories: participation and enhancement. The former involves developing basic communications capabilities like email, maintaining an active website, and allowing passive document sharing. Enhancement involves adopting technologies that alter internal operations or lead to new services. They found that most firms, around 90 percent, made some sort of technology investment. However, only a small fraction (12 percent) adopted Internet technologies that fell into the enhancement category. So while most firms adopted Internet technologies, only a few made investments that would fundamentally change their business.6 2.3. Who Buys Online? We use data from the 2005 Forrester Research Technographics survey, a representative survey of North Americans that asks about respondents’ attitudes toward and use of technology, to form an image of what online shoppers look like. We first look at who uses the Internet in any regular capacity (not necessarily to shop). We run a probit regression of an indicator for Internet use by the respondent during the previous year on a number of demographic variables. The estimated marginal effects are in Table 8.2, column 1. By the time of the 2005 survey, more than 75 percent of the sample reported being online, so the results do not simply reflect the attributes of a small number of technologically savvy early adopters. Internet users are higher-income, more educated, and younger than non-Internet users. The coefficients on the indicators for the survey's household income categories imply that having annual income below $20,000 is associated with a 22 percentage point smaller probability of being online than being in a household with an income over $125,000, the excluded group in the regression. Internet use increases monotonically with income until the $70,000–90,000 range. Additional income seems to have little role in explaining Internet use after that threshold. Education is a sizeable determinant of who is online, even controlling for income. Relative to having a high school degree (the excluded category), not having graduated from high school reduces the probability of using the Page 4 of 28 Online versus Offline Competition Internet by 8 to 9 percentage points (we include categorical education variables for both the (p. 194) (p. 195) female and male household heads), while having a college degree raises it by 6 to 8 points. Table 8.2 Demographics and Probability of Using Internet and Purchasing Online Respondent's race is Black Respondent's race is Asian Respondent's race is other Respondent is Hispanic Respondent is Male Household income 〈 $20K $20K 〈 household income ≤ $30K $30K 〈 household income ≤ $50K $50K 〈 household income ≤ $70K $70K 〈 household income ≤ $90K $90K 〈 household income ≤ $125K Female head of household's education is less than high school Page 5 of 28 Use Internet Purchase Online −0.039 −0.106 (0.007)*** (0.009)*** 0.029 0.01 (0.014)** (0.018) 0.005 −0.002 (0.015) (0.020) 0.005 −0.002 (0.009) (0.012) −0.005 −0.009 (0.005) (0.006) −0.217 −0.309 (0.015)*** (0.011)*** −0.134 −0.207 (0.013)*** 0.012)*** −0.085 −0.133 (0.011)*** (0.011)*** −0.043 −0.085 (0.011)*** (0.011)*** −0.004 −0.038 (0.010) (0.011)*** −0.017 −0.043 (0.010) (0.011)*** −0.081 −0.109 Online versus Offline Competition (0.009)*** (0.012)*** 0.063 0.083 (0.005)*** 0.006)*** −0.091 −0.134 (0.008)*** (0.010)*** 0.084 0.109 (0.004)*** (0.006)*** −0.004 −0.003 (0.001)*** (0.001)** −0.025 −0.085 (0.007)*** (0.011)*** Additional income and family structure controls X X Fraction of sample responding yes 0.763 0.509 N 54,320 54,320 Pseudo-R2 0.24 0.196 Female head of household's education is college Male head of household's education is less than high school Male head of household's education is college Age Age2 /1000 Notes: This table shows the estimates from probit regressions of indicators for households using the Internet and making purchases online on household demographics. The sample includes U.S. households in the 2005 Forrester Research Technographics survey. Standard errors in parentheses. *** p〈0.01, (**) p〈0.05, * p〈0.10. Not surprisingly, the propensity to be online declines with age. The coefficient on the square of age is negative and significant, so the marginal effect grows slightly with age. For example, a 35-year-old is 5.5 percentage points less likely to be online than a 25-year-old, while a 60-year-old is 6.8 percentage points less likely than a 50-year-old to use the Internet. Race also explains some variation in Internet use, though the size of the marginal effect is modest. Blacks are about 4 percentage points less likely to be online than Whites, while Asians are three percentage points more likely. Hispanics are online at the same rate as whites. Gender does not seem to be a factor in explaining Internet use. The results in column 2 of Table 8.2 look at online purchasing behavior per se. The column shows the marginal effects of a probit regression on whether the survey respondent reported making an online purchase within the last Page 6 of 28 Online versus Offline Competition year. The qualitative patterns estimated are similar to those for the probit on Internet use, though many marginal effects have larger magnitudes. So while a low income person (household income less than $20,000 per year) is about 22 percentage points less likely to be online than someone from a household making $125,000 or more, they are 31 percentage points less likely to actually buy something online. Similarly, not having a high school diploma reduces the probability of online purchases by 11 to 13 percentage points relative to having a diploma (as opposed to an 8 to 9 percentage point effect on Internet use), and having a college degree now raises it by 8 to 11 percentage points (as opposed to 6 to 8 points for use). Age effects are also larger, now being in the 8 to 13 percentage point range per 10 years, depending on the ages being compared, as the magnitude of the age effect is still convex. While Blacks were 4 percentage points less likely to be online, they are about 11 percentage points less likely to make purchases once online. On the other hand, while Asians were more likely to be online than Whites and Hispanics, they are not significantly more likely to report having bought goods or services online. (p. 196) Though not shown, we also ran regressions of purchases conditional on Internet use. The results are very similar to the coefficients from the second column. This indicates that selection based on who uses the Internet is not driving the patterns of who purchases products online. These results are informative and largely in line with what we suspect are many readers’ priors. But they reflect overall online purchasing likelihoods, not the determinants of whether consumers, when buying a particular product, choose to do so via online or offline channels. However, the Technographics survey collects additional information on the method of purchase for specific types of products. We can make such comparisons in this case. We investigate consumers’ behavior regarding a set of financial products: auto loans, credit cards, mortgages and home equity loans, auto and life insurance, and checking accounts. The survey asks both whether each of these products were researched online or offline prior to purchase, and whether any purchase was made online or offline. Table 8.3 reports the results. Column 1 of Table 8.3 simply reprints, for the sake of comparison, the results from column 2 of Table 8.2 regarding whether the respondent made any purchase online within the past 12 months. Columns 2 and 3 of Table 8.3 report analogous results for probits on whether the respondent bought any of the particular financial products listed above online within the past year. The results in column 2 are not conditional on the respondent having reported that they researched such financial products online; those in column 3 use the subsample of respondents reporting having researched those product types online. The results from both samples are similar. Many of the qualitative patterns seen for online purchases in general are observed for financial products in particular, but there are some interesting differences. The effect of age is still negative, but is now concave in magnitude rather than convex. And while having a college degree is associated with a significantly higher probability of buying something online, it has a much smaller and insignificant (and in the case of the female head of household, negative) role in financial products. Most striking are the results on race. While Blacks are 11 percentage points less likely to purchase products online than Whites, they are 1.5 percentage points more likely to buy financial products online. Not only is this effect in the opposite direction of the overall results, it is almost as large in magnitude in relative terms.7 Asian and Hispanic respondents are similarly more likely (economically and statistically) to buy financial products online than Whites, even though they did not exhibit statistically different patterns for overall online purchases. We speculate this differential racial pattern for financial products may reflect minorities’ concerns about discrimination in financial product markets, but in the absence of additional evidence, we cannot really know. Finally, we look at changes in consumers’ propensity to buy specific products online in Table 8.4. The second column of the table lists, for a number of product categories that we can follow in the Forrester Technographics survey from 2002 (p. 197) (p. 198) to 2007, the five-year growth rate in consumers’ reported frequency of buying the product online. The third column shows for reference the fraction of consumers reporting having bought the product online in the past year. Auto insurance, one of the financial products we just discussed, saw the fastest growth in online purchases, nearly tripling between 2002 and 2007 (though from an initially small level). Many of the “traditional” online products (if there is such a thing after only about 15 years of existence of e-commerce)— books, computer hardware, airline tickets, and so on—saw more modest but still substantial growth.8 While the growth rate of online purchases for a product is negatively correlated with its 2002 level, the correlation is modest (ρ = -0.13) and not significantly different from zero. Thus it is not the case that the fastest growing products were those that had the slowest start. Page 7 of 28 Online versus Offline Competition Table 8.3 Probability of Purchasing Financial Products Online Financial Products Respondent's race is Black Respondent's race is Asian Respondent's race is other Respondent is Hispanic Respondent is Male Household income 〈 $20K $20K 〈 household income ≤ $30K $30K 〈 household income ≤ $50K $50K 〈 household income ≤ $70K $70K 〈 household income ≤ $90K $90K 〈 household income ≤ $125K Female head of household's education is less than Any product Unconditional Conditional on purchase −0.106 0.015 0.058 (0.009)*** (0.004)*** (0.013)*** 0.01 0.035 0.107 (0.018) (0.009)*** (0.023)*** −0.002 0.001 0.019 (0.020) (0.008) (0.023) −0.002 0.015 0.054 (0.012) (0.006)** (0.016)*** −0.009 0.015 0.033 (0.006) (0.003)*** 0.008)*** −0.309 −0.048 0.107 (0.011)*** (0.004)*** (0.014)*** −0.207 −0.026 0.062 (0.012)*** (0.005)*** (0.015)*** −0.133 −0.017 0.049 (0.011)*** (0.005)*** (0.014)*** −0.085 −0.012 0.035 (0.011)*** (0.005)*** (0.013)*** −0.038 −0.003 0.008 (0.011)*** (0.005) (0.013) −0.043 −0.006 0.016 (0.011)*** (0.005) (0.013) −0.109 −0.011 0.014 Page 8 of 28 Online versus Offline Competition Female head of household's education is less than high school −0.109 −0.011 0.014 (0.012)*** (0.005)** (0.016) 0.083 −0.005 0.013 (0.006)*** (0.003)* 0.008) −0.134 −0.021 0.051 (0.010)*** (0.004)*** (0.012)*** 0.109 0.003 0.008 (0.006)*** (0.003) (0.008) −0.003 −0.005 0.006 (0.001)** (0.001)*** (0.001)*** −0.085 0.015 −0.006 (0.011)*** (0.005)*** (0.014) Fraction of sample responding yes 1 0 0.265 N 54,320 59,173 21,474 Pseudo-R2 0.196 0.097 0.086 Female head of household's education is college Male head of household's education is less than high school Male head of household's education is college Age Age2 /1000 Notes: Estimates from probit regressions of household purchase indicators on demographics. Column 1 reprints for comparison column 2 of Table 8.2. Columns 2 and 3 use an indicator for whether the household purchased one or more of a set of financial products (see text for list) online in the previous year. Column 2 uses the entire sample; column 3 conditions on the subsample that reports having researched financial products online. The sample includes U.S. households in the 2005 Forrester Research Technographics survey. Standard errors in parentheses. (***) p 〈 0.01, (**) p 〈 0.05, (*) p 〈 0.10. 3. How is the Online Channel Different from the Offline Channel? E-commerce technology can affect both demand and supply fundamentals of markets. On the demand side, ecommerce precludes potential customers from inspecting goods prior to purchase. Further, online sellers tend to be newer firms and may have less brand or reputation capital to signal or bond quality. These factors can create information asymmetries between buyers and sellers not present in offline purchases. Online sales also often involve a delay between purchase and (p. 199) (p. 200) consumption when a product must be physically delivered. At the same time, however, e-commerce technologies reduce consumer search costs, making it easier Page 9 of 28 Online versus Offline Competition to (virtually) compare different producers’ products and prices. On the supply side, e-commerce enables new distribution technologies that can reduce supply chain costs, improve service, or both. Both the reduction in consumer search costs and the new distribution technologies combine to change the geography of markets; space can matter less online. Finally, and further combining both sides of the market, online sales face different tax treatment than offline sales. We discuss each of these factors in turn in this section. Table 8.4 Changes in Consumers’ Propensity to Buy Products Online, 2002–2007 Product category Pct. growth in online purchase frequency, 2002–2007 Fraction buying product online, 2007 Car insurance 183.7 0.076 Major appliances 139.6 0.014 Consumer electronics 125.7 0.092 Video games 117.3 0.070 Sporting goods 100.8 0.068 Footwear 89.8 0.116 Credit card 77.2 0.102 Apparel 73.6 0.253 Auto parts 64.3 0.039 Books 60.3 0.278 DVDs 58.6 0.148 Event tickets 53.2 0.121 Music 48.3 0.156 Computer hardware 43.0 0.076 Life insurance 42.2 0.019 Toys 41.2 0.124 Hotel reservations 31.1 0.151 Clothing accessories 23.6 0.089 Airline tickets 22.2 0.172 Page 10 of 28 Online versus Offline Competition Tools/hardware 21.0 0.045 Office supplies 19.1 0.077 Software 12.7 0.113 Flowers 11.0 0.097 Car loans 6.3 0.024 Car rentals 6.2 0.077 Food/beverages 1.1 0.041 Home equity loans 3.5 0.018 Mortgages 25.4 0.025 Small appliances 32.8 0.022 Notes: The table reports both levels of and changes in the fraction of households reporting purchasing specific goods and services online. The sample includes U.S. households in the 2002 and 2007 Forrester Research Technographics surveys. 3.1. Asymmetric Information Information asymmetries are larger when purchasing online for a few reasons. The most obvious is that the consumer does not have the opportunity to physically examine the good at the point of purchase. This presents a potential “lemons problem” where unobservably inferior varieties are selected into the online market. Another is that because online retailing is relatively new, retailers have less brand capital than established traditional retailers. A related factor involves some consumers’ concerns about the security of online transactions. Because information asymmetries can lead to market inefficiencies, both buyers and sellers (particularly sellers of high quality goods) have incentives to structure transactions and form market institutions to alleviate “lemons” problems. Many examples of such efforts on the part of online sellers exist. Firms such as Zappos offer free shipping on purchases and returns, which moves closer to making purchases conditional upon inspection. However, the delay between ordering and consumption inherent to online commerce (more on this below) still creates a wedge. An alternative approach is to convey prior to purchase the information that would be gleaned by inspecting the product. Garicano and Kaplan (2001) examine used cars sold via an online auction, Autodaq, and physical auctions. They find little evidence of adverse selection or other informational asymmetries. They attribute this to actions that Autodaq has taken in order to reduce information asymmetries. Besides offering extensive information on each car's attributes and condition, something that the tools of e-commerce actually make easier, Autodaq brokers arrangements between potential buyers and third-party inspection services. Jin and Kato (2007) examine the market for collectable baseball cards and describe how the use of third-party certification has alleviated information asymmetries. They find a large increase in the use of professional grading services when eBay began being used for buying and selling baseball cards. Another form of disclosure is highlighted in Lewis (2009). Using data from eBay Motors, he finds a positive correlation between the number of pictures that the seller posts and the winning price of the auction. However, he does not find evidence that information voluntarily disclosed by the seller affects the probability that the auction listing results in a sale. Page 11 of 28 Online versus Offline Competition (p. 201) Instead of telling consumers about the product itself, firms can try to establish a reputation for quality or some other brand capital. Smith and Brynjolfsson (2001) use data from an online price comparison site to study the online book market. They find that brand has a significant effect on consumer demand. Consumers are willing to pay an extra $1.72 (the typical item price in the sample is about $50) to purchase from one of the big three online book retailers: Amazon, Barnes & Noble, or Borders. There is evidence that the premium is due to perceived reliability of the quality of bundled services, and shipping times in particular. In online auction markets, rating systems allow even small sellers to build reputations, although Bajari and Hortaçsu (2004) conclude that the evidence about whether a premium accrues to sellers with high ratings is ambiguous. Perhaps a cleaner metric of the effect of reputation in such markets comes from the field experiment conducted by Resnick et al. (2006). There, an experienced eBay seller with a very good feedback rating sold matched lots of postcards. A randomized subset of the lots was sold by the experienced eBay seller, using its own identity. The other subset was sold by the same seller, but using a new eBay identity without any buyer feedback history. The lots sold using the experienced seller identity received winning bids that were approximately eight percent higher. More recently, Adams, Hosken, and Newberry (2011) evaluate whether seller ratings affect how much buyers are willing to pay for Corvettes on eBay Motors. Most of the previous research had dealt with items of small value where the role of reputation might have a relatively modest influence. Collectable sports cars, however, are clearly high value items. In that market, Adams et al. find very little (even negative) effect from seller ratings. In another recent paper, Cabral and Hortaçsu (2010) use a different approach and find an important role for eBay's seller reputation mechanism. They first run crosssectional regressions of prices on seller ratings and obtain results similar to Resnick et al. (2006). Next, using a panel of sellers to examine reputation effects over time, they find that sellers’ first negative feed-back drops their average sales growth rates from +5 percent to –8 percent. Further, subsequent negative feedback arrives more quickly, and the seller becomes more likely to exit as her rating falls. Outside of online auction markets, Waldfogel and Chen (2006) look at the interaction of branding online and information about the company from a third party. They find that the rise of information intermediaries such as BizRate leads to lower market shares for major branded online sellers like Amazon. Thus other sources of online information may be a useful substitute for branding in some markets. 3.2. Delay Between Purchase and Consumption While a lot of digital media that is purchased online can be used/consumed almost immediately after purchase (assuming download times are not a factor), online purchases of physical goods typically involve delivery lags that can range from hours to days and occasionally longer. Furthermore, these delayed-consumption (p. 202) items are the kind of product most likely to be available in both online and brick-and-mortar stores, so the role of this lag can be particularly salient when considering the interaction between a market's online and offline channels. The traditional view of a delay between choice and consumption is as a waiting cost. This may be modeled as a simple discounted future utility flow or as a discrete cost (e.g., Loginova 2009). In either case, this reduces the expected utility from purchasing the good's online version. However, more behavioral explanations hold out the possibility that, for some goods at least, the delay actually confers benefits to the buyer in the form of anticipation of a pleasant consumption experience (e.g., Loewenstein 1987). This holds out the possibility that the impact of delay on the relative advantage of online channels is ambiguous. Though one might think that if delay confers a consistent advantage, offline sellers could offer their consumers the option to delay consumption after purchase rather easily. This, to say the least, is rarely seen in practice. 3.3. Reduced Consumer Search Costs It is generally accepted that search costs online are lower than in offline markets. The rise of consumer information sites, from price aggregation and comparison sites (aka shopbots) to product review and discussion forums, has led to large decreases in consumers’ costs of gathering information. This has important implications for market outcomes like prices, market shares, and profitability, as discussed in detail in Section 4. Online search isn’t completely free; several papers have estimated positive but modest costs. Bajari and Hortaçsu (2003), for example, find the implied price of entering an eBay auction to be $3.20. Brynjolfsson, Dick, and Smith (2010) estimate that the maximum cost of viewing additional pages of search results on a books shopbot is $6.45. Page 12 of 28 Online versus Offline Competition Hong and Shum (2006) estimate the median consumer search cost for textbooks to be less than $3.00. Nevertheless, while positive, these costs are less for most consumers than the value of the time it would take them to travel to just one offline seller. 3.4. Lower Distribution Costs E-commerce affects how goods get from producers to consumers. In some industries, the Internet has caused disintermediation, a diminishment or sometimes the entire removal of links of the supply chain. For example, between 1997 and 2007, the number of travel agency offices fell by about half, from 29,500 to 15,700. This was accompanied by a large increase in consumers’ propensity to directly make travel arrangements—and buy airline tickets in particular—using online technologies.9 E-commerce technologies have also brought changes in how sellers fulfill orders. Firms can quickly assess the state of demand for their products and turn this (p. 203) information into orders sent to upstream wholesalers and manufacturers. This has reduced the need for inventory holding. Retail inventory-to-sales ratios have dropped from around 1.65 in 1992 to 1.34 in late 2010, and from 1.55 to 1.25 over the same period for “total business,” a sum of the manufacturing, wholesale, and retail sectors.10 An example of how increased speed of communication along the supply chain affects distribution costs is a practice referred to as “drop-shipping.” In drop-shipping, retailers transfer orders to wholesalers who then ship directly to the consumer, bypassing the need for a retailer to physically handle the goods. This reduces distribution costs. Online-only retailers in particular can have a minimal physical footprint when using dropshipping; they only need a virtual storefront to inform customers and take orders.11 Randall, Netessine, and Rudi (2006) study the determinants of supply chain choice. Markets where retailers are more likely to adopt drop-shipping have greater product variety, a higher ratio of retailers to wholesalers, and products that are large or heavy relative to their value. Product variety creates a motive for drop-shipping because unexpected idiosyncrasies in variety-specific demand make it costly to maintain the correct inventory mix at the retail level. It is easier to allow a wholesaler with a larger inventory to assume and diversify over some of this inventory risk.12 Similar reasoning shows that drop-shipping is more advantageous when there is a high retailer to wholesaler ratio. Relatively large or heavy products are more likely to be drop-shipped because the higher costs of physically distributing such goods raises the savings from skipping the extra step of shipping from wholesaler to retailer. The Internet has also affected the catalog of products available to consumers. Bricks-and-mortar operations are limited in the number of varieties they offer for sale at one time, as margins from low-volume varieties cannot cover the fixed costs of storing them before sale. Online sellers, however, can aggregate demand for these low-volume varieties over a larger geographic market (this will be discussed in Section 3.5 below). At the same time, they typically have a lower fixed cost structure. The combination of these technological changes lets them offer a greater variety of products for sale. (E-commerce's consumer search tools can also make it easier for consumers of niche products to find sellers.) This “longtail” phenomenon has been studied by Brynjolfsson, Hu, and Smith (2003) and others. Brynjolfsson et al. find that the online book retailers offer 23 times as many titles as did a typical bricks-and-mortar firm like Barnes & Noble. They estimate that this greater product variety generates consumer welfare gains that are 7 to 10 times larger than the gains from increased competition. 3.5. The Geography of Markets E-commerce allows buyers to browse across potential online sellers more easily than is possible across offline outlets. This fading of markets’ geographic boundaries is tied to the reduction in search costs in online channels. Further, e-commerce (p. 204) technologies can reduce the costs of distributing products across wide geographies. The practice of drop-shipping discussed above is an example; not having to ship to retailers can make it easier for supply chains to service greater geo-graphic markets. There is some empirical support for this “death of distance” notion (Cairncross, 1997). Kolko (2000) finds that people in more isolated cities are more likely to use the Internet. Similarly, Sinai and Waldfogel (2004) show that conditional on the amount of local online content, people in smaller cities are more likely to connect to the Internet Page 13 of 28 Online versus Offline Competition than those in larger cities. Forman, Goldfarb, and Greenstein (2005) document that on the margin, businesses in rural areas are more likely to adopt technologies that aid communication across establishments. Despite this, several studies suggest spatial factors still matter. Hortaçsu et al. (2009) look at data from two Internet auction websites, eBay and MercadoLibre. They find that the volume of exchanges decreases with distance. Buyers and sellers that live in the same city have particular preference for trading with one another instead of someone outside the metropolitan area. Hortaçsu et al. surmise that cultural factors and a lower cost of enforcing the contract should a breach occur, explain this result. Blum and Goldfarb (2006) find geography matters online even for purely digital goods like downloadable music, pictures, and movies, where transport and other similar trade costs are nil. They attribute this to culturally correlated tastes among producers and consumers living in relative proximity. Sinai and Waldfogel (2004) find patterns consistent with broader complementarities between the Internet and cities. They find in Media Metrix and Current Population Survey data that larger cities have substantially more local content online than smaller cities, and this content leads people to connect to the Internet.13 We test whether geography matters online more generally by comparing the locations of pure-play online retailers to where people who purchase products online live. If e-commerce makes geography irrelevant, we would expect the two to be uncorrelated. On the other hand, if online sellers are physically located near customers, this suggests that geography still plays a role in these markets. Unfortunately we cannot distinguish with such data whether the relevant channel is shipping costs, contract enforceability, or something else. We measure the number of online-only businesses in geographic markets using County Business Patterns data on the number of establishments in NAICS industry 45411, “Electronic Shopping and Mail-Order Houses.” This industry classification excludes retailers with any physical presence, even if they are a hybrid operation with an online component. Hence these businesses sell exclusively at a distance. (Though they may not necessarily be online, as they could be exclusively a mail order operation. We consider the implications of this below.) We use the Technographics survey discussed above to compute the fraction of respondents in a geographic market who report making online purchases in the previous year. Our geographic market definition is based on the Component Economic Areas (CEAs) constructed by the U.S. Bureau of Economic Analysis. CEAs are groups of economically connected counties; in many cases, they are a metro area plus some additional outlying counties. (p. 205) There are approximately 350 CEAs in the U.S. (Goldmanis et al. (2010) use the same variable to measure the intensity of local online shopping.) We combine these data sets into an annual panel spanning 1998 to 2007. Table 8.5 shows the results from regressing the number of pure-play online sellers on the fraction of consumers in the local market that purchase products online. We include market fixed effects in the regression because unobserved factors might cause certain markets to be amenable to both online sellers and online buyers. For example, Silicon Valley's human capital is both desired by online retailers and makes Valley consumers apt to shop online.14 We further include logged total employment in the CEA in the regression to control for overall economic growth in the market, and we add year fixed effects to remove aggregate trends. The estimate in the first numerical column of Table 8.5 indicates that as the fraction of consumers purchasing products online in a market increases by ten percentage points, on average another 2.2 electronic shopping and mail-order businesses open in the local area. (Establishment counts have been scaled by 10 for better resolution of some of the correlations by business size category.) While NAICS 45411 can include mail-order businesses that do not sell online, it is seems likely (p. 206) that growth in pure mail-order operations within a market would either be uncorrelated or perhaps even negatively correlated with the growth of online shopping in the market. Hence it is likely that the estimated coefficient reflects growth in the number of pure-play online retailers in response to greater use of e-commerce by local consumers. Page 14 of 28 Online versus Offline Competition Table 8.5 Relationship between Fraction Purchasing Products Online and Number of Online Firms within Local Markets Total online only businesses Online only businesses of given size 1–4 5–9 10–19 20–49 50–99 100+ [1] [2] [3] [4] [5] [6] [7] Fraction purchasing online in market 22.29*** (6.190) 14.87** (4.535) 3.714** (1.234) 2.136** (0.719) 1.318* (0.669) 0.211 (0.315) 0.049 (0.345) Year FEs x x x x x x x Market FEs x x x x x x x Mean of dependent variable 39.16 23.31 6.73 4.24 2.65 0.94 1.29 R2 0.963 0.947 0.942 0.941 0.914 0.856 0.920 N 3378 3378 3378 3378 3378 3378 3378 Notes: The table shows the results from regressing the number of online sales businesses (NAICS 45411, Electronic Shopping and Mail-Order Houses) in a geographic market area (see text for definition) on the fraction of consumers in that area that report making online purchases. The sample is an annual panel constructed using establishment counts from U.S. County Business Patterns data and online purchase data from Forrester Research's Technographics Survey. All specifications also include logged total market employment in the year and market fixed effects. Standard errors clustered at the CEA level are given in parentheses. Standard errors in parentheses. (***) p〈 0.01, (**) p 〈 0.05, (*) p 〈 0.10. The next six columns of Table 8.5 report results from similar regressions that use as the dependent variable counts of NAICS 45411 establishments in various employment size categories. A given increase in online shopping is tied to a larger increase in the number of smaller establishments than bigger ones. If we instead use the natural log of the number of establishments as the dependent variable, the estimated effects are much more uniform across the size distribution of firms. So in percentage terms, increasing the fraction of consumers who shop online in an area proportionally increases the number of firms of all sizes. (Some of the growth of the number of larger establishments may well reflect existing businesses becoming larger rather than de novo entry.) 3.6. Tax Treatment One advantage that many online transactions enjoy over transactions in a physical store is the absence of sales tax. Legally, U.S. citizens are obligated to pay their state's sales or use taxes on their online purchases. This rarely happens in practice, as reporting and payment is left completely to the consumer. Only when the online seller “has nexus” in the consumer's state is the sales tax automatically added to the transaction price by the firm.15 This unevenness of the application of sales taxes could lead to a strong advantage for online retail purchases. For example, consumers in Chicago buying online at the beginning of 2012 would avoid the applicable sales tax of 9.50 percent, a considerable savings. Page 15 of 28 Online versus Offline Competition Goolsbee (2000) provides the empirical evidence on this subject. He uses the Forrester Technographics survey to estimate that the elasticity of the probability of consumers buying products on the Internet with respect to the local tax rate is about 3.5. This estimate implies substantial sensitivity of online purchases to tax treatment. If the average sales tax in his data (6.6 percent) were applied to all online transactions, the number of people purchasing products online would fall by 24 percent. While Goolsbee (2000) estimates the effect of sales tax on the extensive margin (whether a person buys anything online), Ellison and Ellison (2009b) estimate the effect of taxes on a measure of total sales that includes both the extensive and intensive margins. Their findings are similar to Goolsbee’s, further bolstering the case that applying sales taxes to Internet purchases could reduce online retail sales by one-quarter. On the supply side, tax structure can distort firm location decisions. Suppose a firm bases its operations in Delaware to take advantage of the state's lax tax laws. If the firm were to create a distribution center in the Midwest to decrease the time (p. 207) it takes to fulfill orders from the Midwest, then it might choose to open the distribution center in the state with relatively few purchasers. A case study of Barnes & Noble (Ghemawat and Baird, 2004) illustrates this point nicely. When Barnes & Noble first created an online business, the online division was almost entirely separate from the brick-and-mortar store. The one shared resource among the online and offline divisions was the company's book buyers. Even though the two divisions shared buyers, the books to be sold on BarnesandNoble.com were sent to a distribution center in Jamesburg, New Jersey. Books bound for traditional brick-and-mortar stores were sent to different warehousing facilities to make it clear which books would not be subject to sales tax. Further, when BarnesandNoble.com went online in May of 1997, the company initially refused to install kiosks with access to the website in stores. They also avoided delivering books ordered online to their physical stores for pick up by customers. It wasn’t until October 2000 that Barnes & Noble, after struggling to compete with Amazon, decided to forego the sales tax benefits it had enjoyed and integrate its online and offline businesses (Ghemawat and Baird, 2006). 4. How E-Commerce Affects Market Outcomes The changes in demand- and supply-side fundamentals that e-commerce brings can foment substantial shifts in market outcomes from their offline-only equilibrium. These include prices, market shares, profitability, and the type of firms operating in the market. 4.1. Prices Perhaps no market outcome has been studied more intensively in the context of online sales activity than prices. Much of the conventional wisdom and some theoretical work (e.g., Bakos, 1997) have focused on the potential for e-commerce to reduce prices. Both reduced consumer search costs and lower distribution costs—two of the fundamental mechanisms described in the previous section—can act to reduce prices in online markets. Lower search costs make firms’ residual demand curves more elastic, reducing their profit-maximizing prices. Reduced distribution costs directly impact profit-maximizing prices if they reflect changes in marginal costs.16 A body of empirical work has supported these predictions about lower prices. For example, Brynjolfsson and Smith (2000) and Clay, Krishnan, and Wolff (2001) find that prices drop due to the introduction of online book markets. Scott Morton, (p. 208) Zettelmeyer, and Silva-Risso (2001) document that consumers who used an online service to help them search for and purchase a car paid on average two percent less than other consumers. Brown and Goolsbee (2002) estimate that price comparison websites led to drops of 8–15 percent in the prices of term life insurance policies. Sengupta and Wiggins (2006) document price reductions in airline tickets driven by online sales. Many of the price reductions documented in these studies and others result from e-commerce technologies making markets more competitive, in the sense that firms’ cross-price elasticities rise. We will discuss below how this can be beneficial for firms with cost advantages over their competitors. However, these same competitive forces can also give strong incentives to firms with cost disadvantages to limit the impact of price differentials. These firms would like to take actions that reduce the propensity of consumers, now with enhanced abilities to shop around, to shift their purchases toward lower-cost sellers. Page 16 of 28 Online versus Offline Competition Certainly, some barriers to substitution exist online. E-commerce markets are not the utterly frictionless commoditytype markets sometimes speculated about early in the Internet's commercial life. Often, more than just the product upon which the transaction is centered is being sold. Goods are usually bundled with ancillary services, and the provision of these services might vary across sellers without being explicitly priced. Sellers’ brands and reputations might serve as a proxy or signal for the quality of such service provision. Smith and Brynjolfsson's (2001) aforementioned study on brand effects in online book sales is an example. Waldfogel and Chen (2006), while finding that price comparison websites weaken brand effects, also find that brand still matters for sellers in a number of product markets. And the work by Jin and Kato (2006), Resnick et al. (2006), and Cabral and Hortaçsu (2010) on seller reputation on online auction sites further bolsters the importance of such ancillary services. Given these results, it is not surprising that firms that operate online—especially those with higher costs than their competitors—try to emphasize brand and bundled services rather than the raw price of the good itself. Ancillary service provision and branding efforts aren’t the only tools firms use to soften price competition. Ellison and Ellison (2009a) document active efforts by online sellers of computer CPUs and memory cards to obfuscate their true prices in order to defeat the price-comparison abilities of e-commerce technologies. In this market, both products and sellers are viewed by consumers as homogeneous, so many sellers focus their efforts on “bait-andswitch”-type tactics where a bare-bones model of the product (often missing key parts most users would find necessary for installation) is priced low to grab top rankings on shopbots, while the additional necessary parts are sold at considerable mark ups. Ellison and Ellison describe a constant battle between sellers trying to find new ways to hide true prices from the shopbots (by making posted prices look very low) and shopbot firms adjusting their information gathering algorithms to better decipher goods’ actual prices. However, Baye and Morgan (2001) make an interesting point about shopbots and other product comparison websites. Building a perfect shopbot—one that (p. 209) reports all information relevant to consumers’ purchasing decisions, allowing them to find their highestutility options almost costlessly—may not be an equilibrium strategy when products are differentiated primarily by price or other vertical attributes. A product comparison site that works too well will destroy the very dispersion in price or other attributes it was created to address, obviating the need for its services. Baye and Morgan show that product comparison websites should provide enough information to be useful for searching customers (on whom the sites rely for revenues, either through subscriptions as in the model or, more often in practice, through advertising revenues), but not so useful as to eliminate their raison d’etre. These active efforts by e-commerce firms are reasons why, as documented by Baye, Morgan, and Scholten (2007) and the studies cited therein, substantial price dispersion remains in most online markets. See chapter 9 in this Handbook for extensive discussion of price comparison sites. 4.2. Other Market Outcomes The advent of online sales in a product market is likely to affect more than just prices. Reduced consumer search costs or differential changes in distribution costs across producers can lead to a wave of creative destruction that shifts the fundamental structure of an industry. Because e-commerce technologies make it easier for consumers to find lower-price sellers, lower-cost firms (or those able to deliver higher quality at the same cost) will grab larger shares of business away from their highercost competitors. Even if, as discussed above, the more competitive landscape created by lower search costs reduces prices and margins, this market structure response could be large enough that low-cost firms actually become more profitable as e-commerce spreads. High-cost firms, on the other hand, are doubly hit. Not only does their pricing power fall, their market share falls too, as customers who were once captive—either because of ignorance or lack of alternatives—flee to better options elsewhere. Some of these firms will be forced out of business altogether. Conventional wisdom suggests that market structure impacts could be large; the rapid growth of online travel sites at the expense of local travel agencies is one oftcited example. But while many academic studies of the effect of ecommerce on prices exist, only a small set of studies have investigated which businesses most benefit and most suffer from e-commerce. Goldmanis et al. (2010) flesh out how such shifts could happen in a model of industry equilibrium where Page 17 of 28 Online versus Offline Competition heterogeneous firms sell to a set of consumers who differ in their search costs. Firm heterogeneity arises from differences in marginal costs, though the model can be easily modified to allow variation in product quality levels instead. Industry consumers search sequentially when deciding from whom to buy. Firms set prices given consumers’ optimal search behavior as well as their (p. 210) own and their rivals’ production costs. Firms that cannot cover their fixed costs exit the industry, and initial entry into the industry is governed by an entry cost. Interpreting the advent and diffusion of e-commerce as a leftward shift in the consumer search cost distribution, Goldmanis et al. show that, consistent with previous literature, opening the market to online sales reduces the average price in the market. The more novel implications regard the equilibrium distribution of firm types, however. Here the model predicts that introducing e-commerce should shrink and sometimes force the exit of low-type (i.e., high-cost) firms and shift market share to high-type (low-cost) firms. Further, new entrants will on average have lower costs than the average incumbent, including those forced out of the market. Testing the model's predictions in three industries perceived to have been considerably impacted by e-commerce —travel agencies, book-stores, and new auto dealers—Goldmanis et al. find support for these predictions. While they cannot measure costs directly in their data, they use size to proxy for firms’ costs. (A considerable body of research has documented that higher cost firms in an industry tend to be smaller. See, e.g., Bartelsman and Doms, 2000.) They find that growth in consumers’ online shopping is linked to drops in the number of small (and presumably high-cost) establishments, but has either no significant impact or even positive impact on the number of the industries’ large establishments. In addition to these industry-wide shifts, e-commerce's effects varied by local markets among bookstores and new car dealers. Cities where consumers’ Internet use grew faster in a particular year saw larger drops (gains) in the number of small (large) bookstores and car dealers over the same year. This also informs the discussion above about whether online sales truly eliminate spatial boundaries in markets.17 The effects among car dealers are particularly noteworthy in that auto manufacturers and dealers in the U.S. are legally prohibited from selling cars online. Therefore any effects of e-commerce must be channeled through consumers’ abilities to comparison shop and find the best local outlet at which to buy their car, not through changes in the technology of car distribution. While this technology-based channel is important in some industries, the consumer-side search channel is the one posited in their model, and therefore new car dealers offer the most verisimilitude to the theory from which they derive their predictions. We add to Goldmanis et al.'s original data and specifications here. Figure 8.1 shows how the composition of employment in the same three industries changed between 1994 and 2007. Each panel shows the estimated fraction of employment in the industry that is accounted for by establishments of three employment size classes: those having 1–9 employees, those with 10–49, and those with 50 or more. In addition to the three industries studied in Goldmanis et al., the figure also shows for the sake of comparison the same breakdown for total employment in the entire County Business Patterns coverage frame (essentially all establishments in the private nonfarm business sector with at least one employee).18 Page 18 of 28 Online versus Offline Competition Click to view larger Figure 8.1 Estimated Share of Industry Employment by Establishment Size. Notes: These figures show the fraction of industry or sector employment accounted for by establishments of differing employment levels. Values are taken from the U.S. County Business Patterns data base (available at http://www.census.gov/econ/cbp/index.html). Industry definitions are as follows: travel agencies, SIC 4724/NAICS 561510; bookstores, SIC 5942/NAICS 451211; and new auto dealers, SIC 5510/NAICS 441110. Panel A shows the breakdown for travel agencies. It is clear that during the early half of the sample period, which saw the introduction and initial diffusion of e-commerce, the share of industry employment accounted for by travel agency offices with (p. 211) (p. 212) fewer than 10 employees shrank considerably. This lost share was almost completely taken up by establishments with 50 or more employees. After 2001, the share losses of the smallest offices stabilized, but the 10–49 employee category began to lose share to the largest establishments. These patterns are consistent with the predictions of the theory—the largest offices in the industry benefit at the cost of the smaller offices. Panel B shows the same results for bookstores. Here, the pattern is qualitatively similar, but even more stark quantitatively. While the fraction of employment at stores with 10–49 employees is roughly stable over the entire period, the largest bookstores gained considerable share at the expense of the smallest. Panel C has the numbers for new car dealers. In this industry, establishments with fewer than 10 employees account for a trivial share of employment, so the interest is in the comparison between the 10–49 employee dealers and those with more than 50. Again, we see that the large establishments accounted for a greater fraction of industry employment over time, with the largest establishments gaining about 10 percentage points of market share at the cost of those with 10–49 employees. Finally, panel D does the same analysis for all establishments in the private nonfarm business sector. It is apparent that the shifts toward larger establishments seen in the three industries of focus were not simply reflecting a broader aggregate phenomenon. Employment shares of establishments in each of the three size categories were stable throughout the period. These predictions about the market share and entry and exit effects of introducing an online sales channel in an industry are based on the assumption that firms behave non-cooperatively. If e-commerce technologies instead Page 19 of 28 Online versus Offline Competition make it easier for firms to collude in certain markets, e-commerce technologies might actually make those markets less competitive. Campbell, Ray, and Muhanna (2005) use a dynamic version of Stahl (1989) to show theoretically that if search costs are high enough initially, e-commerce-driven reductions in search costs can actually make it easier for collusion to be sustained in equilibrium, as they increase the profit difference between the industry's collusive and punishment (static Nash Equilibrium) states. A more direct mechanism through which online sales channels support collusion is that the very transparency that makes it easier for consumers to compare products can also make it easier for colluding firms to monitor each other's behavior. This makes cheating harder. Albæk, Møllgaard, and Overgaard (1997) document an interesting example of this, albeit one that doesn’t directly involve online channels, in the Danish ready-mixed concrete industry. In 1993, the Danish antitrust authority began requiring concrete firms to regularly publish and circulate their transactions prices. Within a year of the institution of this policy, prices increased 15–20 percent in absence of any notable increases in raw materials costs or downstream construction activity. The policy—one that, ironically, was implemented with hopes of increasing competition—facilitated collusion by making it easier for industry firms to coordinate on anticompetitive prices and monitor collusive activities. Online markets are often characterized by easy access to firms’ prices. If it is hard for firms to offer secret discounts because of market convention, technological constraints, or legal strictures, this easy access fosters a powerful monitoring device for colluders. (p. 213) 5. Implications of Online Commerce for Firm Strategy The fundamental effects of opening a concurrent online sales channel in an industry that we discussed in Section 3 can have implications for firms’ competitive strategies. These strategy choices can in turn induce and interact with the equilibrium changes we discussed in Section 4. This section reviews some of these strategic factors. A key factor—perhaps the key factor—influencing firms’ joint strategies toward offline and online markets is the degree of connectedness between online and offline markets for the same product. This connectedness can be multidimensional. It can involve the demand side: how closely consumers view the two channels as substitutes. It can involve the supply side: whether online and offline distribution technologies are complementary. And it can involve firms’ available strategy spaces: how much leeway firms have in conducting separate strategic trajectories across channels, which is particularly salient as it regards how synchronized a firm's pricing must be across offline and online channels. At one extreme would be a market where the offline and online channels are totally separated. Specifically, consumers view the product as completely different depending upon the channel through which it is sold (perhaps there are even separate online and offline customer bases); there are no technological complementarities between the two channels; and firms can freely vary positioning, advertising, and pricing of the same product across the channels. In this case, each channel can be thought of as an independent market. The firm's choices in each channel can be analyzed independently, as there is no scope for strategic behavior that relies upon the interplay between the two channels. Of more interest to us here—and where the research literature has had to break new ground—are cases where there are nontrivial interactions between online and offline channels selling the same products. We’ll discuss some of the work done in this area below, categorizing it by the device through which the online and offline channels are linked: consumer demand (e.g., substitutability), technological complementarities, or strategic restrictions. 5.1. Online and Offline Channels Linked Through Consumer Demand One way the online and offline sales channels can be connected is in the substitutability that buyers perceive between the channels. The extent of such substitutability determines two related effects of opening an online channel in a market: the potential for new entrants into an online channel to steal away business from incumbents, and the amount of cannibalization offline incumbents will suffer upon opening an online segment. Not all consumers in a market will necessarily (p. 214) view this substitutability symmetrically. Distinct segments can react differently to the presence of online purchase options. The observed substitutability simply reflects the aggregate impact of these segments’ individual responses. Page 20 of 28 Online versus Offline Competition These factors have been discussed in several guises in the literature investigating the strategic implications of operating in a market with both online and off-line channels. Dinlersoz and Pereira (2007), Koças and Bohlmann (2008), and Loginova (2009) construct models where heterogeneity in consumers’ views toward the substitutability of products sold in the two segments affects firms’ optimal strategies. Dinlersoz and Pereira (2007) and Koças and Bohlmann (2008) build models where some customers have loyalty for particular firms and others buy from the lowest-price firm they encounter. Offline firms with large loyal segments (“Loyals”) stand to lose more revenue by lowering their prices to compete in the online market for price-sensitive “Switchers.” Hence the willingness of incumbents from the offline segment to enter new online markets depends in part on the ratios of Loyals-to Switchers. This also means the success of pure-play online firms is tied to the number of Switchers. In some circumstances, opening an online channel can lead to higher prices in the offline market, as the only remaining consumers are Loyals who do not perceive the online option as a substitute. Depending on the relative valuations and sizes of the Loyals and Switchers segments, it is even possible that the quantity-weighted average price in the market increases. In effect, the online channel becomes a price discrimination device. Direct tests of these models are difficult, but they do imply that if we compare two firms, the one with the higher price will have more loyal consumers than the other. We can conduct a rough test of this in the bookselling industry using the Forrester Technographics data. In it, consumers are asked whether they have shopped either online or offline at Amazon, Barnes & Noble, or Borders in the previous thirty days. Clay et al. (2002) found that Amazon set prices higher than Barnes & Noble, which in turn set prices higher than Borders. Thus the models predict that Amazon's customers will be more loyal than Barnes & Noble’s, who are themselves more loyal than Borders’. In our test, this implies that of the customers of these sellers, Amazon will have the highest fraction of exclusive shoppers, followed by Barnes & Noble and Borders. The results are in Table 8.6. In the first row, the first column reports the fraction of consumers who purchased a book in the past three months and shopped only at Amazon. The second column gives the fraction of customers who purchased a product from Amazon as well as from Barnes & Noble or Borders. If we take the first column as a crude measure of the fraction of Amazon's loyal customers and the second column as a measure of those willing to shop around, Amazon's customer base is roughly split between Loyals and Switchers. While the models would predict that, given the observed price difference, Barnes and Noble's Loyals-to-Switchers ratio should be lower, this is not the case in the data, as reflected in the second row. However, Borders’ low ratio of Loyals to Switchers is consistent with them having the lowest prices. A caveat to these results, however, is that they could be confounded by Internet use. The models’ predictions regard the loyalty of (p. 215) a firm's online customers. If many of Barnes & Noble's loyal customers are offline, our measure might overstate the loyalty of Barnes & Noble's online consumers. We address this in the second panel of Table 8.6 by recalculating the fractions after conditioning on the consumer having purchased a book online. Now the evidence is exactly in line with the predictions of Dinlersoz and Pereira (2007) and Koças and Bohlmann (2008): the rank ordering of the firms’ prices is the same as the ordering of the Loyals-to-Switchers ratio. In Loginova (2009), consumers’ ignorance of their valuations for a good shapes the nature of the link between online and offline markets. Consumers in her model differ in their valuations for the market good, but do not realize their valuations until they either (a) visit an offline retailer and inspect the good, or (b) purchase the good from an online retailer (no returns are allowed). Under certain parameter restrictions, there is an equilibrium where both channels are active and all consumers go to offline retailers and learn their valuations. Upon realizing their utility from the good, they decide either to immediately purchase the good from the offline retailer or to go home and purchase the product from an online retailer while incurring a waiting cost. This creates an equilibrium market segmentation where consumers with low valuations buy from online stores and high-valuation consumers buy immediately at the offline outlet they visited. The segmentation lets offline retailers raise their prices above what they would be in a market without an online segment. The imperfect substitutability between online and offline goods segments the market and allows firms to avoid head-on competition. These papers focus on the extent to which goods sold on online and offline channels are substitutes, but it is possible in certain settings that they may be complements. Empirical evidence on this issue is relatively sparse. Gentzkow (2007) (p. 216) estimates whether the online edition of the Washington Post is a substitute or complement for the print edition. The most basic patterns in the data suggest they are complements: consumers Page 21 of 28 Online versus Offline Competition who visited the paper's website within the last five days are more likely to have also read the print version. However, this cross sectional pattern is confounded by variation in individuals’ valuations from consuming news. It could be that some individuals like to read a lot of media, and they often happen to read the online and offline versions of the paper within a few days of one another. But conditioning on having read one version, that specific individual may be less likely to read the other version. This is borne out in a more careful look at the data; instrumenting for whether the consumer has recently visited the paper's website using shifters of the consumer's costs of reading online, Gentzkow finds the two channels’ versions are rather strong substitutes. Using a different methodology, Biyalogorsky and Naik (2003) look at whether Tower Records’ introduction of an online channel lifted or cannibalized its offline sales. They find cannibalization, though it was modest, on the order of 3 percent of the firm's offline sales. Given that brick-and-mortar record stores have clearly suffered from online competition since this study, their result suggests that much of the devastation was sourced in across-firm substitution rather than within-firm cannibalization. Table 8.6 “Switchers” and “Loyals” in the Book Industry Consumers Who Purchased Books in Past Three Months Loyals Switchers Loyals/Switchers Amazon 0.201 0.203 0.990 Barnes & Noble 0.279 0.278 1.004 Borders 0.087 0.153 0.569 Consumers Who Purchased Books Online in Past Three Months Loyals Switchers Loyals/Switchers Amazon 0.343 0.262 1.309 Barnes & Noble 0.179 0.274 0.653 Borders 0.034 0.095 0.358 Notes: Entries under “Loyals” are the fraction of customers who purchased from only one of the three firms listed, while “Switchers” are the fraction purchasing from more than one of the three firms. The third column gives the ratio of Loyals to Switchers for each firm. The top panel includes all consumers who purchased books, whether online or offline, while the lower panel only includes consumers who purchased books online. Data are from Forrester Research's Technographics Survey. 5.2. Online and Offline Channels Linked Through Technological Complementarities Wang (2007) ties the online and offline channels with a general complementarity in the profit function that he interprets as a technological complementarity. His model treats the introduction of e-commerce into an industry as the opening of a new market segment with lower entry costs. The model's dynamic predictions are as follows. Taking advantage of the new, lower entry costs, pure-play online sellers enter first to compete with the brick-andmortar incumbents. But the complementarity between the online sales and distribution technology and the offline technology gives offline incumbents incentive to expand into the online channel. It also gives these firms an inherent advantage in the online market, as they are able to leverage their offline assets to their gain. As a result, many of the original online-only entrants are pushed out of the industry. Thus a hump-shaped pattern is predicted in the number of pure-play online firms in a product market, and a steady diffusion of former offline firms into the online channel. Page 22 of 28 Online versus Offline Competition This is a reasonably accurate sketch of the trajectory of the online sector of many retail and service markets. The online leaders were often pure-play sellers: Amazon, E-Trade, Hotmail, pets.com, and boo.com, for example. But many of these online leaders either eventually exited the market or were subsumed by what were once offline incumbents. Some pure-play firms still exist, and a few are fabulously successful franchises, but at the same time, many former brick-and-mortar sellers now dominate the online channels of their product markets. (p. 217) Jones (2010) explores a different potential complementarity. The notion is that the online technology is not just a way to sell product, but it can also be an information gathering tool. Specifically, the wealth of data generated from online sales could help firms market certain products to individuals much more efficiently and lead to increased sales in both channels. 5.3. Online and Offline Channels Linked Through Restrictions on Strategy Space Liu, Gupta, and Zhang (2006) and Viswanathan (2005) investigate cases where the online and offline channels are tied together by restrictions on firms’ strategy spaces—specifically, that their prices in the two channels must be a constant multiple of one another. In the former study, this multiple is one: the firm must price the same whether selling online or offline. Viswanathan (2005) imposes that the price ratio must be a constant multiple, though not necessarily one. While it might seem unusual that these pricing constraints are exogenously imposed instead of arising as equilibrium outcomes, it is true that certain retailers have faced public relations and sometimes even legal problems due to differences in the prices they charge on their websites and in their stores. Liu, Gupta, and Zhang remark that many multichannel firms report in surveys that they price consistently across their offline and online channels. Liu, Gupta, and Zhang (2006) show that when the equal pricing restriction holds, an incumbent offline seller can deter the entry of a pure-play online retailer by not entering the online market itself. This seemingly counterintuitive result comes from the uniform price requirement across channels. An incumbent moving into the online channel is restricted in its ability to compete on price, because any competition-driven price decrease in the online market reduces what the incumbent earns on its inframarginal offline units. This limit to its strategy space can actually weaken the incumbent's competitive response so much that a pure-play online retailer would be more profitable if the incumbent enters the online segment (and therefore has to compete head-to-head with one hand tied behind its back) than if the incumbent stays exclusively offline. Realizing this, the incumbent can sometimes deter entry by the pure-play online firm by staying out of the online channel in the first place. The link across the online and offline channels in this model creates an interesting situation in which the offline firm does not gain an advantage by being the first mover in to the online channel. Instead, it may want to abstain from the online market altogether. Viswanathan (2005) models the online and offline models as adjacent spatial markets. Consumers in one market cannot buy from a firm in the other market. However, one firm at the junction of the two markets is allowed to operate as a dual-channel supplier, but it must maintain an exogenously given price ratio of k between the two markets. Viswanathan shows that in this setup, the price charged (p. 218) by the two-channel firm will be lower than the offline-only firms’ prices but higher than the pure-play online sellers. 6. Conclusions The emergence of online channels in a market can bring substantial changes to the market's economic fundamentals and, through these changes, affect outcomes at both the market level and for individual firms. The potential for such shifts has implications in turn for firms’ competitive strategies. Incumbent offline sellers and new pure-play online entrants alike must account for the many ways a market's offline and online channels interact when making pricing, investment, entry, and other critical decisions. We have explored several facets of these interactions in this chapter. We stress that this is only a cursory overview, however. Research investigating these offline-online connections is already substantial and is still growing. This is rightly so, in our opinion; we expect the insights drawn from this literature to only become more salient in the future. Online channels have yet to fully establish themselves in some markets and, in those where they have been developed, are typically growing faster than bricks-and-mortar channels. This growing salience is especially likely in the retail and services sectors, where online sales appear to still have substantial room for Page 23 of 28 Online versus Offline Competition growth. Acknowledgment We thank Martin Peitz and Joel Waldfogel for comments. Syverson thanks the NSF and the Stigler Center and Centel Foundation/Robert P. Reuss Faculty Research Fund at the University of Chicago Booth School of Business for financial support. References Adams, C.P., Hosken, L., Newberry, P., 2011. Vettes and Lemons on eBay. Quantitative Marketing and Economics 9(2), pp. 109–127. Albæk, S., Møllgaard, P., Overgaard, P.B., 1997. Government-Assisted Oligopoly Coordination? A Concrete Case. Journal of Industrial Economics 45(4), pp. 429–443. Bajari, P., Hortaçsu, A., 2003. The Winner's Curse, Reserve Prices, and Endogenous Entry: Empirical Insights from eBay Auctions. RAND Journal of Economics 34(2), pp.329–355. Bajari, P., Hortaçsu, A., 2004. Economic Insights from Internet Auctions. Journal of Economic Literature 42(2), pp. 257–286. Bakos, J.Y., 1997. Reducing Buyer Search Costs: Implications for Electronic Marketplaces. Management Science 43(12), pp. 1676–1692. Bartelsman, E.J., Doms, M., 2000. Understanding Productivity: Lessons from Longitudinal Microdata. Journal of Economic Literature 38(3), pp.569–594. Baye, M. R., Morgan, J., 2001. Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets. American Economic Review 91(3), pp.454–474. (p. 221) Baye, M. R., Morgan, J., Scholten, P., 2007. Information, Search, and Price Dispersion. Handbooks in Economics and Information Systems, vol. 1, (T. Hendershott, Ed.), Amsterdam and Boston: Elsevier. pp. 323–376. Biyalogorsky, E., Naik, P., 2003. Clicks and Mortar: The Effect of On-line Activities on Of-line Sales. Marketing Letters 14(1), pp. 21–32. Blum, B.S., Goldfarb, A., 2006. Does the Internet Defy the Law of Gravity? Journal of International Economics, 70(2), pp. 384–405. Brown, J.R., Goolsbee, A., 2002. Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry. Journal of Political Economy 110(3), pp. 481–507. Brynjolfsson, E., Dick, A.A., Smith, M.D., 2010. A Nearly Perfect Market? Differentiation vs. Price in Consumer Choice. Quantitative Marketing and Economics 8(1), pp. 1–33. Brynjolfsson, E., Smith, M.D., 2000. Frictionless Commerce? A Comparison of Internet and Conventional Retailers. Management Science 46(4), pp. 563–585. Brynjolfsson, E., Hu, Y., Smith, M.D., 2003. Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety at Online Booksellers. Management Science 49(11), pp. 1580–1596. Cabral, L., Hortaçsu, A., 2010. The Dynamics of Seller Reputation: Evidence from eBay. Journal of Industrial Economics 58(1), pp. 54–78. Cairncross, F., 1997. The Death of Distance: How the Communication Revolution Will Change Our Lives. Harvard Business School Press. Page 24 of 28 Online versus Offline Competition Campbell, C., Ray, G., Muhanna, W.A., 2005. Search and Collusion in Electronic Markets. Management Science 51(3), pp. 497–507. Clay, K., Krishnan, R., Wolff, E., 2001. Prices and Price Dispersion on the Web: Evidence from the Online Book Industry. Journal of Industrial Economics 49(4), pp. 521–539. Clay, K., Krishnan, R., Wolff, E., 2002. Retail Strategies on the Web: Price and Non-price Competition in the Online Book Industry. Journal of Industrial Economics 50(3), pp. 351–367. Dinlersoz, E.M., Pereira, P., 2007. On the Diffusion of Electronic Commerce. International Journal of Industrial Organization 25(3), pp. 541–574. Ellison, G., Ellison, S.F., 2009a. Search, Obfuscation, and Price Elasticities on the Internet. Econometrica 77(2), pp. 427–452. Ellison, G., Ellison, S.F., 2009b. Tax Sensitivity and Home State Preferences in Internet Purchasing. American Economic Journal: Economic Policy 1 (2), pp.53–71. Forman, C., Goldfarb, A., Greenstein, S., 2003. Which Industries Use the Internet? Organizing the New Industrial Economy, vol. 12 (Advances in Applied Microeconomics), (M. Baye, Ed.), Elsevier. pp. 47–72. Forman, C., Goldfarb, A., Greenstein, S., 2005. How did Location Affect Adoption of the Commercial Internet? Global Village vs. Urban Leadership. Journal of Urban Economics 58, pp. 389–420. Garicano, L., Kaplan, S.N., 2001. The Effects of Business-to-Business E-Commerce on Transaction Costs. Journal of Industrial Economics 49(4), pp. 463–485. Gentzkow, M., 2007. Valuing New Goods in a Model with Complementarity: Online Newspapers. American Economic Review 97(3), pp. 713–744. Ghemawat, P., Baird, B., 2004. Leadership Online (A): Barnes & Noble vs. Amazon.com. Boston, Mass: Harvard Business School Publishing. (p. 222) Ghemawat, P., Baird, B., 2006. Leadership Online (B): Barnes & Noble vs. Amazon.com in 2005. Boston, Mass: Harvard Business School Publishing. Goldmanis, M., Hortaçsu, A., Syverson, C., Emre, O., 2010. E-commerce and the Market Structure of Retail Industries. Economic Journal 120(545), pp. 651–682. Gollop, F.M., Monahan, J.L., 1991. A Generalized Index of Diversification: Trends in U.S. Manufacturing. Review of Economics and Statistics 73(2), pp. 318–330. Goolsbee, A., 2000. In a World without Borders: The Impact of Taxes on Internet Commerce. Quarterly Journal of Economics 115(2), pp. 561–576. Hong, H., Shum, M., 2006. Using Price Distributions to Estimate Search Costs. The RAND Journal of Economics 37(2), pp. 257–275. Hortaçsu, A., Martinez-Jerez, F.A., Douglas, J., 2009. The Geography of Trade in Online Transactions: Evidence from eBay and MercadoLibre. American Economic Journal: Microeconomics 1(1), pp. 53–74. Jin, G. Z., Kato, A., 2006. Price, Quality, and Reputation: Evidence from an Online Field Experiment. RAND Journal of Economics 37(4), pp. 983–1005. Jin, G. Z., Kato, A., 2007. Dividing Online and Offline: A Case Study. Review of Economic Studies 74(3), pp. 981– 1004. Jones, S.M., 2010. Internet Poised to Become Bigger Force in Retail. Chicago Tribune. Accessed January 9, 2010 at www.chicagotribune.com/business/chi-tc-biz-outlook-retail-0105-jan06,0,6844636.story. Koças, C., Bohlmann, J.D., 2008. Segmented Switchers and Retailer Pricing Strategies. Journal of Marketing 72(3), Page 25 of 28 Online versus Offline Competition pp. 124–142. Kolko, J., 2000. The Death of Cities? The Death of Distance? Evidence from the Geography of Commercial Internet Usage. The Internet Upheaval (Ingo Vogelsang and Benjamin M. Compaine, eds.), Cambridge: MIT Press. pp. 73–98. Krishnan, K., Rao, V., 1965. Inventory Control in N Warehouses. Journal of Industrial Engineering 16, pp. 212–215. Lewis, G., 2009. Asymmetric Information, Adverse Selection and Online Disclosure: The Case of eBay Motors. American Economic Review 101(4), pp. 1535–1546. Liu, Y., Gupta, S., Zhang, Z.J., 2006. Note on Self-Restraint as an Online Entry-Deterrence Strategy. Management Science 52(11), pp.1799–1809. Loewenstein, G., 1987. Anticipation and the Valuation of Delayed Consumption. Economic Journal 97(387), pp. 666–684. Loginova, O., 2009. Real and Virtual Competition. Journal of Industrial Economics 57(2), pp.319–342. Mesenbourg, T., 2001. Measuring Electronic Business: Definitions, Underlying Concepts, and Measurement Plans. 〈www.census.gov/epcd/www/ebusines.htm〉 Morton, F.S., Zettelmeyer, F., Silva-Risso, J., 2001. Internet Car Retailing. Journal of Industrial Economics 49(4), pp.501–519. Mukhopadhyay, T., Kekre, S., Kalathur, S., 1995. Business Value of Information Technology: A Study of Electronic Data Interchange. MIS Quarterly 19(2), pp. 137–156. Netessine, S., Rudi, N., 2006. Supply Chain Choice on the Internet. Management Science 52(6), pp. 844–864. Randall, T., Netessine, S., Rudi, N., 2006. An Empirical Examination of the Decision to Invest in Fulfillment Capabilities: A Study of Internet Retailers. Management Science 52(4), pp. 567–580. Resnick, P., Zeckhauser, R., Swanson, J., Lockwood, K., 2006. The Value of Reputation on eBay: A Controlled Experiment. Experimental Economics 9(2), pp. 79–101. (p. 223) Saloner, G, Spence, A.M., 2002. Creating and Capturing Value—Perspectives and Cases on Electronic Commerce, Crawfordsville: John Wiley & Sons, Inc. Sengupta, A., Wiggins, S.N., 2006. Airline Pricing, Price Dispersion and Ticket Characteristics On and Off the Internet. NET Institute Working Paper, No. 06–07. Sinai, T., Waldfogel, J., 2004. Geography and the Internet: Is the Internet a Substitute or a Complement for Cities? Journal of Urban Economics, 56(1), pp. 1–24. Smith, M.D., Brynjolfsson, E., 2001. Consumer Decision-Making at an Internet Shopbot: Brand Still Matters. Journal of Industrial Economics 49(4), pp.541–558. Stahl, D.O., II, 1989. Oligopolistic Pricing with Sequential Consumer Search. American Economic Review 79(4), pp. 700–712. US Census Bureau, 2010. 2008 E-commerce Multi-Sector Report. 〈www.census.gov/estats . Viswanathan, S., 2005. Competing across Technology-Differentiated Channels: The Impact of Network Externalities and Switching Costs. Management Science 51(3), pp. 483–496. Waldfogel, J., Chen, L., 2006. Does Information Undermine Brand? Information Intermediary Use and Preference for Branded Web Retailers. Journal of Industrial Economics 54(4), pp. 425–449. Wang, Z., 2007. Technological Innovation and Market Turbulence: The Dot-Com Experience. Review of Economic Dynamics 10(1), pp.78–105. Page 26 of 28 Online versus Offline Competition Notes: (1.) Ghemawat and Baird (2004, 2006) offer a detailed exploration of the nature of competition between Amazon and Barnes & Noble. (2.) The Census Bureau defines e-commerce as “any transaction completed over a computer-mediated network that involves the transfer of ownership or rights to use goods or services.” A “network” can include open networks like the internet or proprietary networks that facilitate data exchange among firms. For a review of how the Census Bureau collects data on e-commerce and the challenges posed in quantifying e-commerce, see Mesenbourg (2001). (3.) The Census Bureau defines the B2B and B2C distinction similarly to the sector-level definition here. It is worth noting, however, that because the Bureau does not generally collect transaction-level information on the identity of the purchaser, these classifications are only approximate. Also, the wholesale sector includes establishments that the Census classifies as manufacturing sales branches and offices. These are locations separate from production facilities through which manufacturers sell their products directly rather than through independent wholesalers. (4.) The Census Bureau tracks retail trade e-commerce numbers at a higher frequency. As of this writing, the latest data available are for the fourth quarter of 2011, when e-commerce-related sales accounted for a seasonallyadjusted 4.8 percent of total retail sales. (5.) The R&D data is aggregated across some of the 3-digit industries, so when comparing online sales shares to R&D, we aggregate the sales channel data to this level as well. This leaves us 17 industries to compare. Additionally, the product differentiation index (taken from Gollop and Monahan 1991) is compiled using the older SIC system, so we can only match 14 industries in this case. (6.) The two-digit NAICS industry with the highest enhancement category investment rate (28 percent) was Management of Companies and Enterprises (NAICS 55). The lowest adoption rate (6.2 percent) was in Educational Services (NAICS 61). (7.) Note that when comparing the magnitudes of the coefficient estimates across columns in Table 8.3, one should be mindful of the average probability of purchase in the sample, pbar, displayed at the bottom of the table. Because the average probability of purchasing one of the financial products online (9.6 percent) is roughly one-fifth the probability that any product is purchased (50.9 percent), the estimated marginal effects in the financial products’ case are five times the relative size. Thus the 1.5-percentage-point marginal effect for Black respondents and financial products in column 2 corresponds to a roughly 7.5-percentage-point marginal effect in column 1. (8.) Two products saw substantial declines in online purchase likelihoods: mortgages and small appliances. The former is almost surely driven by the decline in demand for mortgages through any channel. We are at a loss to explain the decline in small appliance purchases. (9.) An interesting case where the Internet brought about increased intermediation is in auto sales. There, at least in the United States, legal restrictions require that all sales go through a physical dealer who cannot be owned by a manufacturer. Given these restrictions, online technologies in the industry were devoted to creating referral services like Autobytel.com. Consumers shop for and select their desired vehicle on the referral service's website, and then the service finds a dealer with that car and has the dealer contact the consumer with a price quote (Saloner and Spence, 2002). (10.) http://www.census.gov/mtis/www/data/text/mtis-ratios.txt, retrieved 1/26/11. (11.) The practice has been adopted by many but not all online-only retailers. Netessine and Rudi (2006) report that 31 percent of pure-play Internet retailers use drop-shipping as their primary method of filling orders. (12.) Traditional retailers have used other mechanisms to serve a similar function (though likely at a higher cost). For example, retailers with multiple stores often geographically pool inventory risk by cross-shipping orders from a store with an item in inventory to one that takes a customer order but is stocked out (e.g., Krishnan and Rao 1965). (13.) As noted above, the same authors find that conditional on local content, people from smaller cities are more Page 27 of 28 Online versus Offline Competition likely to connect to the Internet. Interestingly, in their data, these two forces are just offset so that use of the Internet isn’t strongly correlated with city size. (14.) We have also estimated specifications that control for the fraction of the local population that uses the Internet for any purpose. (This variable is similarly constructed from the Technographics survey.) This did not substantively impact the nature of the results described below, except to make the estimated positive effect of online shopping on online retailers larger. (15.) The great majority of states have a sales tax; only Alaska, Delaware, Montana, New Hampshire, and Oregon do not. Whether a firm has nexus within a state is not always obvious. In the Supreme Court decision Quill vs. North Dakota (1992), it was established that online merchants without a substantial physical presence in the state would not have to enforce sales tax in that state. Later, the 1998 Internet Tax Nondiscrimination Act clarifies that a web presence in a state does not constitute nexus. (16.) Asymmetric information can affect prices as well, though the direction of this effect is ambiguous. Quantities, however, should decline if information becomes more asymmetric. (17.) The aggregate impact observed among travel agencies resulted from the nature of the institutional shifts in industry revenues that e-commerce caused. Responding to a shift in customers toward buying tickets online, airlines cut ticket commissions to travel agents, which accounted for 60 percent of industry revenue in 1995, completely to zero by 2002. These commission cuts were across the board, and did not depend on the propensity of travelers to buy tickets online in the agents’ local markets. (18.) County Business Patterns do not break out actual total employment by size category, so we impute it by multiplying the number of industry establishments in an employment category by the midpoint of that category's lower and upper bounds. For the largest (unbounded) size categories, we estimated travel agency offices and bookstores with 100 or more employees had an average of 125 employees; auto dealers with more than 250 employees had 300 employees. Imputations were not necessary in the case of the total nonfarm business sector, as the CBP do contain actual employment by size category in that case. Ethan Lieber Ethan Lieber is a PhD student in the Economics Department at the University of Chicago. Chad Syverson Chad Syverson is Professor of Economics at the Booth School of Business, University of Chicago. Page 28 of 28 Comparison Sites Oxford Handbooks Online Comparison Sites Jose-Luis Moraga-Gonzalez and Matthijs R. Wildenbeest The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0009 Abstract and Keywords This article, which reviews the work on comparison sites, discusses how comparison sites operate and their main economic roles. It also reports a model of a comparison site. The important issue of price discrimination across channels is explored. It is noted that the comparison site becomes a marketplace more attractive for the buyers than the search market. The market clears when transactions between firms and consumers take place. The fact that firms can price discriminate eliminates the surplus consumers obtain by opting out of the price comparison site and this ultimately destroys the profits of the retailers. The analysis of the simple model of a comparison site has exhibited that product differentiation and price discrimination play a critical role. Click-through data can assist the assessment of structural models of demand. Keywords: comparison sites, price discrimination, marketplace, firms, consumers, product differentiation, click-through data 1. Introduction Not so long ago individuals used atlases, books, magazines, newspapers, and encyclopedias to find content. To locate businesses and their products, it was customary to use yellow pages, directories, newspapers, and advertisements. Family and friends were also a relevant source of information. Nowadays things are quite different: many individuals interested in content, a good, or a service usually conduct a first search on the Internet. Through the Internet individuals can easily access an immense amount of information. Handling such a vast amount of information has become a complex task. Internet browsers constitute a first tool to ease the navigation experience. Thanks to the browsers, users move easily across documents and reach a large amount of information in just a few mouse-clicks. Search technologies are a second tool to facilitate the browsing experience. They are fundamental to navigate the Internet because they help users locate and aggregate content closely related to what they are interested in. Search engines such as Google, Yahoo, and Bing constitute one type of search technologies. These search tools are often offered at no cost for users and the placing of advertisements constitutes the most important way through which these search engines are financed. By taking advantage of the keywords the user provides while searching for information, search engines can pick and deliver targeted advertisements to the most interested audience. In this way, search engines become a relatively precise channel through which producers and retailers can reach consumers. This raises the search engines’ value and their scope (p. 225) to extract rents from producers or retailers. The study of the business model of search engines constitutes a fascinating research area in economics (see e.g. Chen and He, 2011; Athey and Ellison, 2011; Spiegler and Eliaz, 2011; and Gomes, 2011). Comparison sites, or shopping robots (shopbots) such as PriceGrabber.com, Shopper.com, Google Product Search, and Bing Shopping are a second type of search technologies. These sites help users find goods or services that are sold online. For multiple online vendors, shopbots provide a significant amount of information, including the products they sell, the prices they charge, indicators about the quality of their services, their delivery costs as well as their payment methods. By using shopbots, consumers can easily compare a large number of alternatives available in the market and ultimately choose the most satisfactory one. Because they collate information from various offers relatively quickly, shopbots reduce consumer search costs considerably. Business models vary across comparison sites. Most shopbots do not charge consumers for access to their sites and therefore the bulk of their profits is obtained via commercial relationships with the shops they list. They get paid via subscription fees, click-through fees, or commission fees. Some comparison sites list sellers at no cost and Page 1 of 21 Comparison Sites get their revenue from sponsored links or sponsored ads. Finally, some charge consumers to obtain access to its information but firms do not pay any fees. The emergence of Internet shopbots in the marketplace raises questions not only about the competitiveness and efficiency of product markets but also about the most profitable business model. Do all types of firms have incentives to be listed in comparison sites? Why are the prices listed in these sites dispersed, even if the advertised products are seemingly homogeneous? Do comparison sites enhance social welfare? How much should each side of the market pay for the services offered by comparison sites? Addressing these questions is the main focus of this chapter. We do this within a framework where a comparison site designs its fee structure to attract (possibly vertically and horizontally differentiated) online retailers, on the one hand, and consumers, on the other hand. While analyzing our model, we describe the received wisdom in some detail. The study of search engines other than comparison sites raises other interesting economic questions. The economics of online advertising is described by Anderson (2012) in chapter 14 of this volume. Of particular importance is the management of sponsored search advertisements. Search engines cannot limit themselves to delivering consumer access to the advertiser placing the highest bid; they must also carefully manage the quality of the ads; otherwise they can lose the ability to obtain surplus from advertisers. The rest of this chapter is organized as follows. In Section 2, we describe how comparison sites operate, as well as their main economic roles. We also summarize the main results obtained in the small theoretical literature in economics, and discuss empirical research on the topic. In Section 3, we present a model of a comparison site. The model is then applied to comparison sites dealing with homogeneous products. Later we discuss the role of product differentiation, both horizontal and (p. 226) vertical. We also explain the important issue of price discrimination across channels. We conclude the chapter with a summary of theoretical and empirical considerations and put forward some ideas for further research. 2. Comparison Sites Comparison sites or shopbots are electronic intermediaries that assist buyers when they search for product and price information in the Internet. Shopbots have been operating on the Internet since the late 1990s. Compared to other more traditional intermediary institutions, most shopbots do not sell items themselves—instead they gather and aggregate price, product, and other relevant information from third-party sellers and present it to the consumers in an accessible way. By doing this, consumers can easily compare offers. Shopbots also display links to the vendors’ websites. These links allow a buyer to quickly navigate to the site of the seller that offers her the best deal. Shopbots operate according to several different business models. The most common is that users can access the comparison site for free, while sellers have to pay a fee. Initially, most comparison sites charged firms a flat fee for the right to be listed. More recently, this fee usually takes the form of a cost-per-click and is paid every time a consumer is referred to the seller's website from the comparison site. Most traditional shopbots, like for instance PriceGrabber.com and Shopping.com, operate in this way. Fees typically depend on product category—current rates at PriceGrabber range from $0.25 per click for clothing to $1.05 per click for plasma televisions. Alternatively, the fee can be based on the execution of a transaction. This is the case of Pricefight.com, which operates according to a cost-per-acquisition model. This model implies that sellers only pay a fee if a consumer buys the product. Other fees may exist for additional services. For example, sellers are often given the possibility to obtain priority positioning in the list after paying an extra fee. A second business model consists of offering product and price comparison services for free to both sellers and buyers and thus relies on advertising as a source of revenue. Both Google Product Search and Microsoft's Bing Shopping are examples of comparison sites that have adopted this type of business model. Any seller can list products in these websites by uploading and maintaining a product data feed containing information about the product price, availability, shipping costs, and so on. A third, although less common, model is to have consumers pay a membership fee to access the comparison site, whereas sellers are listed for free. AngiesList.com for instance aggregates consumer reviews about local service companies, which can be accessed by consumers for an annual membership fee between $10 and $50, depending on where the consumer lives. (p. 227) 2.1. Early Intermediation Literature Shopbots are platforms through which buyers and sellers can establish contact with one another. In this sense, comparison sites essentially play an intermediation role. As a result, we are first led to the literature on intermediation, which has been a topic of interest in economics in general, and in finance in particular. Spulber (1999), in his study of the economic role and relevance of intermediaries, describes various value-adding roles played by intermediaries. The following aspects are prominent: buyer and seller aggregation, lowering of search and matching costs, and facilitation of pricing and clearing Page 2 of 21 Comparison Sites services. In a market where buyers and sellers meet and negotiate over the terms of trade, a number of business risks and costs exist. Reducing such risks and costs is a key role played by intermediaries. In terms of forgone welfare opportunities, not finding a suitable counter-party is in itself the most costly hazard trading partners face; sometimes a trading partner is found at a cost but either rationing occurs or the failure to reach a satisfactory agreement takes place, in which case similar welfare losses are realized. In all these situations, an intermediary can enter the market and reduce the inefficiencies. By announcing prices publicly, and by committing to serve orders immediately, intermediaries reduce significantly the costs of transacting. Intermediaries “make the market” by choosing input and output prices to maximize their profits. Market makers trade on their own account and so they are ready to buy and sell in the market in which they operate. Provision of immediacy, which is a key aspect emphasized in the seminal articles of Demsetz (1968) and Rubinstein and Wolinsky (1987), distinguishes market makers from other intermediating agents in the value chain. An example of a market maker is a supermarket chain with sufficient upstream bargaining power so as to have an influence on bid and ask prices. In finance, perhaps the most common examples of market-makers are stock exchange specialists (such as commercial banks, investment financial institutions, and brokers). Market makers are also the subject of study in Gehrig (1993), Yavas (1996), Stahl (1988), and Spulber (1999). Watanabe (2010) extends the analysis by making the intermediation institution endogenous. Rust and Hall (2003) distinguish between market-makers that post take-it-or-leave-it prices and middlemen who operate in the over-the-counter market at prices that can be negotiated. They study the conditions under which brokers and market-makers can coexist, and study the welfare properties of intermediated equilibria. Price comparison sites are similar to traditional intermediaries in that they “facilitate” trade between online shoppers and retailers. However, what distinguishes a comparison site from a traditional intermediary is that the latter typically buys goods or services from upstream producers or sellers and resells them to consumers. Shopbots do not trade goods, but add value by aggregating information. In that sense, shopbots are more similar to employment agencies and realtors, who also serve the purpose of establishing a bridge between the supply and the demand sides of the market. (p. 228) How can a price comparison site enter the market and survive in the long run? Do comparison sites increase the competitiveness of product markets? Do they enhance market efficiency? This chapter revolves around these three questions. Whether a comparison site can stay in business in the long run is not, a priori, clear. The problem is that, given that retailers and consumers can encounter each other outside the platform and conduct transactions, the search market constitutes a feasible outside option for the agents. In fact, a comparison site can only stay in business if it chooses its intermediation fees carefully enough to out-compete the search market, or at least to make intermediated search as attractive as the search market. The question is then whether a comparison site can indeed create value for retailers and consumers. The first paper studying these questions is Yavas (1994). Yavas studies the match-making role of a monopolistic intermediary in a competitive environment. He shows that the intermediary can obtain a profit by attracting high-valuation sellers and lowvaluation buyers; the rest of the agents trade in the decentralized market. Interestingly, relative to the market without intermediary, buyers and sellers lower their search intensity, which can ultimately decrease matching rates and cause a social welfare loss. Though Yavas’ analysis is compelling, most markets are populated by firms that hold a significant amount of market power. Since market power drives a wedge between the market outcome and the social optimum, it cannot be ignored when modeling the interaction between comparison sites, retailers, and consumers in real-world markets. Our work especially adds in this direction. 2.2. Our Model and Results, and Their Relation to the Literature Our model, described in detail in Section 3, aims at understanding how comparison sites can overcome “local” market power and emerge in the marketplace. In addition, we study whether comparison sites enhance social welfare. “Local” market power can stem from geographical considerations, from loyalty, or from behavioral-type of assumptions such as random-shopping or default-bias (Spiegler, 2011). Our model is inspired from Baye and Morgan's (2001) seminal paper. Baye and Morgan had geographical considerations in mind when they developed their model so in their case “local” market power arises from geographical market segmentation. From a general point of view, however, the particular source of “local” market power is not very important. We will assume buyers opting out of the comparison site will buy at random, and therefore this will be the main source of “local” market power. In essence, the model we study is as follows. Suppose that in a market initially characterized by some sort of segmentation, a price comparison site is opened up. Suppose that the comparison site initially succeeds at attracting some of the buyers from the various consumer segments. The comparison site creates value for firms (p. 229) since a firm that advertises its product on the comparison site can access consumers “located” in other segmented markets. This is reinforcing in that consumers, by registering with the shopbot, can observe a number of product offerings from the advertising firms in addition to the usual one. We study the extent to which the market becomes centralized. We also compare the levels of welfare attained with and without a comparison site. It turns out that product differentiation, both vertical and horizontal, and Page 3 of 21 Comparison Sites the possibility to price discriminate between the centralized and the decentralized marketplaces play an important role. We describe next the results we obtain and how they connect with earlier work. We first study the case in which retailers sell homogeneous products. This is the case examined in Baye and Morgan (2001). We show that a crucial issue is whether online retailers can practice price discrimination across marketplaces or not. If price discrimination is not possible, as in Baye and Morgan, then the platform's manager has an incentive to raise firms’ participation fees above zero so as to induce less than complete participation of the retailers. This results in an equilibrium with price dispersion, which enhances the gains users obtain from registering with the shopbot. Although establishing a price comparison site is welfare improving, the equilibrium is not efficient because prices are above marginal costs and the comparison site only attracts a share of the transactions. We show that the market outcome is quite different when retailers can price discriminate across marketplaces, that is, when they are allowed to advertise on the price comparison site a price different from the one they charge in their websites. In that case, the utility consumers derive from buying randomly, which is the outside option of consumers, is significantly reduced and the price comparison site can choose its tariffs such that both consumers and firms fully participate, while still extracting all rents from the market. This means that with price discrimination the market allocation is efficient and all trade is centralized. We then move to study the case in which retailers sell horizontally differentiated products. This case was examined by Galeotti and Moraga-González (2009). We employ the random utility framework that gives rise to logit demands. In such a model, we show that the price comparison site can choose fees that fully internalize the network externalities present in the market. These fees attract all firms to the platform, which maximizes the quality of the matches consumers obtain and thereby the overall economic rents. The market allocation is not fully efficient because product sellers have market power. However, the monopolist intermediary does not introduce distortions over and above those arising from the market power of the differentiated product sellers. The fact that the comparison site attracts all retailers and buyers to the platform does not depend on whether the retailers can price discriminate across marketplaces or not. This result stems from the aggregation role played by the (product and price) comparison site. By luring firms into the platform not only price competition is fostered so consumers benefit from lower prices but also more choice is offered to consumers. Since the comparison site becomes an aggregator of variety, the comparison site becomes a marketplace more attractive for the buyers than the search market. (p. 230) In our final model, we allow for vertical product differentiation in addition to horizontal product differentiation. The main result we obtain is that the nature of the pricing policy of the comparison site can change significantly and produce an inefficient outcome. Note that when quality differences across retailers are absent, the comparison site obtains the bulk of its profits from the buyers. By lowering the fees charged to the firms, more value is created at the platform for consumers and this value is in turn extracted by the comparison site via consumer fees. We show that when quality differences are large, the comparison site may find it profitable to do otherwise by charging firm fees sufficiently high so as to prevent low-quality producers from participating in the comparison site. This raises the rents of the high-quality sellers, and at the same time creates value for consumers. These rents are in turn extracted by the comparison site via firm and consumer participation fees. In this equilibrium, the intermediary produces a market allocation that is inefficient. 2.3. Empirical Literature Empirical studies centered around shopbots have focused on distinct issues. A number of these studies look at whether predictions derived from the theoretical comparison site models are in line with the data. Using micro data on individual insurance policies, Brown and Goolsbee (2002) provide empirical evidence that increased usage of comparison sites significantly reduced the price of term life insurance in the 1990s, while prices did not fall with increased Internet usage in the period before these comparison sites began. Brynjolfsson and Smith (2001) use click-through data to analyze behavior of consumers searching for books on Dealtime.com. They find that shopbot consumers put substantial brand value on the biggest three retailers (Amazon, Barnes and Noble, and Borders), which suggests it is indeed important to model product differentiation. Baye, Morgan, and Scholten (2004) analyze more than four million price observations from Shopper.com and find that price dispersion is quite persistent in spite of the increased usage of comparison sites. Baye, Gatti, Kattuman, and Morgan (2006) look at how the introduction of the Euro affected prices and price dispersion using data from Kelkoo, a large comparison site in the European Union. They find price patterns broadly consistent with predictions from comparison site models. More recently, Moraga-González and Wildenbeest (2008) estimate a model of search using price data for memory chips obtained from the comparison site MySimon.com. Their estimates can be interpreted so as to suggest that consumer participation rates are relatively low – between 4 and 13 percent of the consumers use the search engine. They find significant price dispersion. An, Baye, Hu, Morgan, and Shum (2010) structurally estimate a model of a comparison site using British data from Kelkoo and use the estimates to simulate the competitive effects of horizontal mergers. (p. 231) Finally, some papers use data from comparison sites to estimate demand models. Ellison and Ellison (2009) study competition between sellers in a market in which the comparison site Pricewatch.com played a dominant role and, using sales Page 4 of 21 Comparison Sites data for one of the retailers, find that demand is tremendously price sensitive for the lowest-quality memory modules. In addition Ellison and Ellison find evidence that sellers are using obfuscation strategies, with less elastic demand for higher quality items as a result. Koulayev (2010) estimates demand and search costs in a discrete choice product differentiation model using click-through data for hotel searches, and finds that search frictions have a significant impact on demand elasticity estimates. 3. A Model of a Comparison Site We study a model of a comparison site in which subscribing consumers can compare the prices charged by the different advertising retailers and the characteristics of their products. A comparison site has therefore the features of a two-sided market. Two-sided markets are characterized by the existence of two groups of agents which derive gains from conducting transactions with one another, and the existence of intermediaries that facilitate these transactions. Exhibitions, employment agencies, videogame platforms, Internet portals, dating agencies, magazines, newspapers and journals are other examples of two-sided markets (see Armstrong, 2006; Caillaud and Jullien, 2003; Evans, 2003; and Rochet and Tirole, 2003).1 In our model the comparison site is controlled by a monopolist. One group of users consists of firms selling products and the other group of users is made of consumers. From the point of view of a firm, participating in the comparison site is a way to exhibit its product and post its price to a potentially larger audience. An individual firm that does not advertise on the comparison site limits itself to selling to those customers who remain outside the platform. Likewise, for a consumer, visiting the platform is a way to learn the characteristics and the prices of all the products of the participating firms. A consumer who does not register with the comparison site can only trade outside the platform. We will assume consumers who opt out of the platform randomly visit a firm.2 The monopoly platform sets participation fees for consumers and firms to attract business to the platform. Traditionally, comparison sites have used fixed advertising fees, or flat fees, paid by firms that participate. For the moment we shall assume other fees, like per-click or per-transaction fees, are equal to zero.3 Let denote the (fixed) fee the platform charges firms for participation. While the platform can charge different fees to firms and consumers, we assume the platform cannot price discriminate among firms by charging them different participation fees. Likewise, let denote the fee charged to consumers for registering with the platform. For simplicity, assume the platform incurs no cost. (p. 232) On the supply side of the market, there are two possibly vertically and horizontally differentiated retailers competing in prices. Let us normalize their unit production cost to zero. A retailer may decide to advertise its product and price on the platform (A) or not to advertise it at all (NA). Advertising may involve certain cost k associated to the feeding of product and price information into the comparison site. For the moment, we will ignore this cost. Let Ei = {A, NA} be the set of advertising strategies available to a firm i. A firm i's participation strategy is then a probability function over the set Ei. We refer to α i as the probability with which a firm i chooses A, while 1 − α i denotes the probability with which such a firm chooses . A firm i's pricing strategy on the platform is a price (distribution) Fi. The firm may charge a different price (distribution), denoted Fio, to the consumers who show up at the shop directly.4 A strategy for firm is thus denoted by σi = {α i, Fi, Fio}, i = 1,2. The strategies of both firms are denoted by σ and the (expected) payoff to a firm i given the strategy profile σ is denoted πi(σ). There is a unit mass of consumers. Consumers can either pick a firm at random and buy there, or else subscribe to the platform, view the offerings of the advertising firms and buy the most attractive one. We assume that consumers are distributed uniformly across firms so each firm receives half of the non-subscribing consumers. Those buyers who choose not to register with the platform visit a retailer at random, say i, and buy there at the price Pio. If they participate in the centralized market, they see all the products available and choose the one that matches them best. To keep things simple, we assume that a consumer who registers with the platform cannot trade outside the platform within the current trading period, even if she finds no suitable product in the platform. Consumer m's willingness to pay for the good sold by firm i is The parameter is εim assumed to be independently and identically double exponentially distributed across consumers and products with zero mean and unit variance and can be interpreted as a match parameter that measures the quality of the match between consumer i and product m. We assume there is an outside option with utility u0 = ε0m. Let G and g denote the cumulative and probability distribution functions of εim, respectively.5 A buyer demands a maximum of one unit. To allow for vertical product differentiation, let Δ ≡ δ1 − δ2 〉 0 be the (ex-ante) quality differential between the two products. Ex-ante, consumers do not know which firm sells which quality; like match values, the quality of a particular firm is only known after consumers visit such firm. Buyers may decide to register with the platform (S) or not at all (NS). The set of consumers’ pure strategies is denoted R = {S, NS}. A consumer's mixed strategy is a probability function over the set R. We refer to μ ∈ [0,1] as the probability with which a consumer registers with the platform. Given all other agents’ strategies, u(μ) denotes the Page 5 of 21 Comparison Sites (expected) utility of a consumer who subscribes with probability μ. (p. 233) The timing of moves is the following. In the first stage, the comparison site chooses the participation fees. In the second stage, firms simultaneously decide on their participation and pricing decisions, while consumers decide whether to register with the platform or not. Firms and consumers that do not enter the platform can only conduct transactions when they match in the decentralized market; likewise, consumers who register with the comparison site can only conduct transactions there. The market clears when transactions between firms and consumers take place. We study subgame perfect equilibria.6 3.1. Homogeneous Products It is convenient to start by assuming that firms sell homogeneous products. Therefore, we assume that εim = 0 for all i,m, including the outside option, and that δi = δj. As a result, uim = δ for all i and m. This is the case analyzed in the seminal contribution of Baye and Morgan (2001).7 In what follows, we will only sketch the procedure to derive a SPE. For details, we refer the reader to the original contribution of Baye and Morgan. Let us proceed backwards and suppose that the participation fees a and s are such that firms and consumers are indifferent between participating in the price comparison site or not. Recall that μ denotes the fraction of participating consumers and α the probability with which a firm advertises its price on the platform. To allow for mixed pricing strategies, refer to F(p) as the advertised price. Consider a firm that does not advertise at the price comparison site. This firm will only sell to a fraction (1 − μ)/2 of nonparticipating consumers and therefore it is optimal for this firm to charge the monopoly price δ. As a result, a firm that does not advertise at the price comparison site obtains a profit: Consider now a firm that decides to charge a price p and advertise it at the price comparison site. This firm will sell to a fraction (1 − μ)/2 of non-participating consumers as well as to the participating consumers if the rival either does not advertise or advertises a higher price. Therefore, the profit this firm will obtain is8 It is easy to see that for advertising fees 0 ≤ a 〈 μδ an equilibrium in pure advertising strategies does not exist. In fact, if the rival firm did not advertise for sure, then firm i would obtain a profit equal to ((1 − μ)δ/2 if it did not advertise either, while if it did advertise a price just below δ, this firm would obtain a profit equal to (1 − μ)/2+ μ)δ − a. Likewise, an equilibrium where firms advertise with (p. 234) probability 1 does not exist either if a 〉 0. If the two firms advertised with probability 1, the price comparison site would resemble a standard Bertrand market so both firms would charge a price equal to the marginal cost and then firms would not obtain sufficient profits to cover the advertising fees. Therefore an equilibrium must have α ∈ (0, 1). To solve for equilibrium, we impose the following indifference conditions: These conditions tell us that (i) a firm must be indifferent between advertising and not advertising (first equality) and also that (ii) a firm that advertises the monopoly price on the platform gets the same profits that a firm that advertises any other price in the support of the price distribution (second equality). Setting the condition and solving for α gives the equilibrium advertising policy of the firms: Using the condition and solving for F(p), gives the pricing policy of the firms: The lower bound of the equilibrium price distribution is found by setting F* (p̱) = 0 and solving for p̱, which gives p̱ = δ − 2α * δ Page 6 of 21 Comparison Sites μ/(1+μ). We now turn to the users’ side of the market. Users take as given firms’ behavior so they believe firms to advertise with probability α * a price drawn from the support of F*. A user who registers with the platform encounters two prices in the price comparison site with probability α *2 , in which case she picks the lowest one; with probability 2α * (1−α *) there is just one price advertised on the platform. The expected utility to a user who registers with the price comparison site is then where . A user who does not register with the price comparison site always buys from one of the firms at random so her utility will be equal to zero if the chosen firm does not advertise, while it will be equal to δ − E[p] otherwise. Therefore, (p. 235) As a result, for given advertising and subscription fees, we may have two types of market equilibria. One is when users participate surely. In that case, users must derive a strictly positive utility from subscribing. The other is when users’ participation rate is less than one. In that case, they must be indifferent between participating and not participating so μ must solve the following equation Whether one equilibrium prevails over the other is a matter of expectations: if firms expect users to be pretty active then it is optimal for consumers to do so, and vice-versa. We now fold the game backwards and consider the stage of the game where the manager of the price comparison site chooses the pair of advertising and subscription fees {a, s} to maximize its profits. The profits of the intermediary are: (1) We now argue that a SPE with partial firm and user participation cannot be sustained. Suppose that at the equilibrium fees a and s, we have a continuation game equilibrium with α *, μ* 〈 1. Then, the participation rate of the users μ* is given by the solution to u(1) = u(0) and the participation rate of the firms α * = 1 −a/δμ*. Table 9.1 shows that the intermediary's profits increase monotonically in s. The idea is that in this continuation equilibrium the elasticity of user demand for participation in the price comparison site is strictly positive. In the putative equilibrium u(1) = u(0). Keeping everything else fixed, an increase in s makes utility (p. 236) outside the platform higher than inside the platform. To restore equilibrium firms must advertise more frequently, which can only occur if the fraction of subscribing consumers increases. So an increase in s is accompanied by an increase in the participation rates of both retailers and users, as it can be seen in the table. This clearly shows than in SPE it must be the case that μ = 1. Page 7 of 21 Comparison Sites Table 9.1 Homogeneous Products: Comparison Site's Profits Increase in Consumer Fee s a s E[p] E[min p1, p2] α μ πi u* Π 0.500 0.010 0.933 0.912 0.170 0.602 0.199 0.011 0.176 0.500 0.020 0.899 0.868 0.238 0.656 0.172 0.024 0.251 0.500 0.030 0.870 0.831 0.289 0.703 0.148 0.038 0.310 0.550 0.030 0.867 0.827 0.281 0.765 0.118 0.037 0.332 0.600 0.030 0.864 0.824 0.274 0.826 0.087 0.037 0.353 0.650 0.030 0.862 0.821 0.268 0.887 0.056 0.037 0.374 0.650 0.040 0.834 0.785 0.307 0.938 0.031 0.051 0.437 0.650 0.050 0.807 0.751 0.342 0.988 0.006 0.066 0.494 0.650 0.053 0.800 0.743 0.350 1.000 0.000 0.070 0.508 Notes: The quality parameter δ is set to 1. Given this result, to maximize the profits of the price comparison site we need to solve the problem Table 9.2 shows the solution of this problem for δ = 1. The profits of the intermediary are maximized when a = 0.426 and s = 0.119 (in bold) so in SPE user participation is maximized but firm participation is not. The intuition for this result is clear. If the firms did participate surely, then the price advertised on the platform would be equal to the marginal cost and the firms would not make any money. In such a case the bulk of the profits of the price comparison site would have to be extracted from the consumers. However, in the absence of price dispersion no consumer would be willing to pay anything to register with the platform. To summarize, the key insights from Baye and Morgan (2001) are that the profits of the price comparison site are maximized when consumers participate with probability one but firms do not, in this way creating price dispersion and so subscription value for consumers. As we have mentioned above, price dispersion is ubiquitous also on the Internet (Baye, Morgan and Scholten, 2004) and, moreover, it is difficult to reject the null hypothesis that advertised prices on comparison sites are random (Moraga-González and Wildenbeest, 2008). Page 8 of 21 Comparison Sites Table 9.2 Homogeneous Products Model: Maximum Profits Intermediary a s E[p] E[min p1, p2] α μ πi u* 0.100 0.140 0.256 0.165 0.900 1.000 0.000 0.670 0.320 0.200 0.162 0.402 0.299 0.800 1.000 0.000 0.478 0.482 0.300 0.151 0.516 0.415 0.700 1.000 0.000 0.339 0.571 0.400 0.127 0.611 0.519 0.600 1.000 0.000 0.233 0.607 0.426 0.119 0.633 0.544 0.574 1.000 0.000 0.210 0.608 0.450 0.111 0.653 0.567 0.550 1.000 0.000 0.191 0.607 0.500 0.097 0.693 0.614 0.500 1.000 0.000 0.153 0.597 0.600 0.066 0.766 0.701 0.400 1.000 0.000 0.094 0.546 0.700 0.040 0.832 0.783 0.300 1.000 0.000 0.050 0.460 Notes: The quality parameter δ is set to 1. (p. 237) 3.1.1. The Role of Price Discrimination Across Marketplaces In Baye and Morgan (2001) price comparison sites have incentives to create price dispersion since by doing so they create value for consumers. If prices on- and off-platform are similar, consumers are not interested in the platform services so the bulk of the money has to be made on the firms’ side. By raising firms participation fees, a price comparison site achieves two objectives at once. On the one hand, competition between firms is weakened and this increases the possibility to extract rents from the firm side; on the other hand, price dispersion goes up and this in turn increases the possibility to extract rents from the user side of the market. This interesting result is intimately related to the assumption that online retailers cannot price discriminate across marketplaces. To see this, suppose that a firm could set a price at its own website which is different from the one advertised on the platform. Since consumers who do not register with the platform are assumed to pick a website at random, it is then obvious that the website's price would be equal to δ. A firm participating in the price comparison site would then obtain a profit equal to In such a case, imposing the indifference conditions for equilibrium that would yield Now we argue that, to maximize profits, the price comparison site wishes to induce full participation of firms and consumers, thereby maximizing social welfare and extracting all the surplus. The expected utility to a user who registers with the price comparison site is given by the same expression for u(1) above, i.e., u(1) = α *2 (δ − E[min{p1, p2 }])+2α * (1−α *) (δ−E[p])−s. A user who does not register with the price comparison site buys at the price δ so her utility will be u(0)=0. Page 9 of 21 Comparison Sites Consider now the stage of the game where the price comparison site's manager chooses firm and consumer participation fees. Suppose that at the equilibrium fees firms and consumers mix between participating and not participating. Table 9.3 shows that the elasticity of consumer participation is also positive in this case. As a result, the intermediary will continue to raise the user fees until all consumers participate with probability one. (p. 238) Table 9.3 Homogeneous Products with Price Discrimination: Intermediary's Profits Increase in s a s E[p] E[min p1, p2] α μ πi u* Π 0.500 0.010 0.948 0.932 0.100 0.556 0.222 0.000 0.106 0.500 0.025 0.916 0.890 0.158 0.594 0.203 0.000 0.173 0.500 0.050 0.879 0.842 0.224 0.644 0.188 0.000 0.256 0.550 0.050 0.879 0.842 0.224 0.708 0.146 0.000 0.281 0.600 0.050 0.879 0.842 0.224 0.773 0.114 0.000 0.307 0.650 0.050 0.879 0.842 0.224 0.837 0.081 0.000 0.333 0.650 0.075 0.849 0.803 0.274 0.895 0.052 0.000 0.423 0.650 0.100 0.822 0.770 0.316 0.951 0.025 0.000 0.506 0.650 0.123 0.800 0.743 0.350 1.000 0.000 0.000 0.508 Notes: The quality parameter δ is set to 1. Table 9.4 Homogeneous Products Model with Price Discrimination: Intermediary's Maximum Profits a s E[p] E[min p1, p2] α μ πi u* Π 0.750 0.063 0.863 0.822 0.250 1.000 0.000 0.000 0.438 0.650 0.123 0.800 0.743 0.350 1.000 0.000 0.000 0.508 0.500 0.250 0.693 0.614 0.500 1.000 0.000 0.000 0.750 0.250 0.563 0.462 0.359 0.750 1.000 0.000 0.000 0.938 0.100 0.810 0.256 0.166 0.900 1.000 0.000 0.000 0.990 0.010 0.980 0.047 0.019 0.990 1.000 0.000 0.000 1.000 0.001 0.998 0.007 0.002 1.000 1.000 0.000 0.000 1.000 0.000 1.000 0.000 0.000 1.000 1.000 0.000 0.000 1.000 Notes: The quality parameter δ is set to 1. In Table 9.4 we show that the price comparison site's profits increase by lowering the firms fee and increasing the users charge until they reach 0 and 1, respectively, which implies that all firms and consumers participate with probability one. In that case, product prices are driven down to marginal cost and the bulk of the profits of the intermediary is obtained from the consumers. Notice that with price discrimination, the market allocation is efficient. The intuition is the following. By increasing competition Page 10 of 21 Comparison Sites in the platform, prices go down and more surplus is generated for consumers. This surplus, since buyers have in practice no outside option, can be extracted from consumers via participation (p. 239) fees. Interestingly, the fact that firms can price discriminate eliminates the surplus consumers obtain by opting out of the price comparison site and this ultimately destroys the profits of the retailers. 3.1.2. Click-Through Fees In recent years many comparison sites have replaced the fixed-fees by a cost-per-click (CPC) tariff structure. In this subsection we show that the CPC business model does not alter the main insights we have obtained so far.9 Let be the clickthrough fee. Assume also that there still exists a (small) fixed cost a firm has to pay for advertising on the platform, denoted k, which can be interpreted as a hassle cost for feeding up the comparison site information system. The profits of a firm that advertises on the platform then become The first part of this profits expression comes from the consumers who do not register with the platform and buy at random. The second part comes from the consumers who participate in the comparison site. These consumes click on firm i's offer when firm i is the only advertising firm, which occurs with probability 1= α, or when firm j also participates at the clearinghouse but advertises a higher price than firm i, which occurs with probability α(1−F(p)). Using the condition and solving for α gives If μ consumers participate in the platform, the total expected number of clicks is then μ(1−(1−α *)2 ). Using the condition and solving for F gives The decision of consumers does not change—they participate with a probability μ*, which is the solution to u(1)=u(0). The profits of the intermediary are then Table 9.5 shows that the property that the elasticity of the consumer demand for participation is positive still holds. As a result, the platform will increase until all consumers register with the price comparison site. (p. 240) Page 11 of 21 Comparison Sites Table 9.5 Homogeneous Products Model with CPC: Platform's Profits Increase in Consumer Fee s k c s E[p] E[min p1, p2] α μ πi u* Π 0.200 0.500 0.010 0.952 0.938 0.255 0.537 0.231 0.012 0.125 0.200 0.500 0.020 0.925 0.903 0.355 0.620 0.190 0.027 0.193 0.200 0.500 0.030 0.899 0.870 0.428 0.700 0.150 0.043 0.257 0.200 0.550 0.030 0.900 0.872 0.436 0.788 0.106 0.044 0.319 0.200 0.600 0.030 0.900 0.873 0.445 0.901 0.050 0.044 0.401 0.200 0.600 0.033 0.892 0.864 0.465 0.935 0.033 0.050 0.431 0.200 0.600 0.036 0.884 0.854 0.484 0.969 0.015 0.056 0.462 0.200 0.600 0.039 0.877 0.846 0.500 1.000 0.000 0.061 0.488 Notes: The quality parameter δ is set to 1. Table 9.6 Homogeneous Products Model with CPC: Intermediary's Maximum Profits k c s E[p] E[min p1, p2] α μ πi u* Π 0.200 0.100 0.145 0.487 0.393 0.778 1.000 0.000 0.399 0.240 0.200 0.200 0.127 0.570 0.487 0.750 1.000 0.000 0.323 0.315 0.200 0.300 0.108 0.651 0.579 0.714 1.000 0.000 0.249 0.383 0.200 0.400 0.086 0.730 0.670 0.667 1.000 0.000 0.180 0.442 0.200 0.500 0.063 0.805 0.759 0.600 1.000 0.000 0.117 0.483 0.200 0.563 0.048 0.851 0.814 0.542 1.000 0.000 0.081 0.493 0.200 0.600 0.039 0.877 0.845 0.500 1.000 0.000 0.061 0.489 0.200 0.700 0.014 0.943 0.927 0.333 1.000 0.000 0.019 0.403 Notes: The quality parameter δ is set to 1. Given this observation, to maximize the profits of the comparison site we need to solve Table 9.6 shows the solution of this problem for δ = 1 and k = 0.2. The profits of the intermediary are maximized when c = 0.563 so again in SPE user participation is maximized but firm participation is not. The intuition for this result is the same as above. (p. 241) 3.2. Differentiated Products We continue with the case in which firms sell products that are both horizontally and vertically differentiated. Galeotti and Moraga-González (2009) study a similar model, but without vertical product differentiation, that is, δi = δ for all i, and with Page 12 of 21 Comparison Sites match values uniformly distributed. Here we explore the role of vertical product differentiation and we choose to work with logit demands for convenience. Later, we shall see the relationship between these two models. Recall that μ is the proportion of consumers in the market registering with the comparison site and that α i is the probability firm i advertises her product at the comparison site. We first look at the price set by a firm that does not advertise at the comparison site. Denoting by pio such a price, it should be chosen to maximize the profits a firm obtains from the consumers who visit a vendor at random. Since consumer m's outside option generates utility u0 = ε0m, we have (2) Let denote the monopoly price of firm i, and the profits this firm obtains if it does not advertise. Taking the first-order condition of equation (2) with respect to pio and solving for pio gives where W[exp(δi − 1)] is the Lambert W-Function evaluated at exp(δi − 1), i.e., the W that solves exp(δi − 1) = W exp(W). Suppose firm i decides to advertise a price pi at the comparison site. The expected profits of this firm are where α j is the probability the rival firm advertises at the comparison site. As can be seen from this equation, demand depends on whether or not the rival firm (p. 242) advertises. With probability α j seller i competes with seller j for the proportion μ of consumers visiting the comparison site, which means that in order to make a sale the utility offered by firm i needs to be larger than the utility offered by firm j as well as the outside good. Since we are assuming εim is i.i.d. type I extreme value distributed this happens with probability exp(δi − pi)/(1+ ∑k=i,j exp(δk − pk)). Similarly, with probability (1−α j) seller i is the only firm competing on the platform, which means ui only has to be larger than u0 for firm i to gain these consumers, i.e., exp(δi−pi)/(1+exp(δi−pi)). Finally, a proportion (1−μ)/2 of consumers buy at random, and the probability of selling to these consumers is exp(δi−pi)/(1+exp(δi−pi)). Page 13 of 21 Comparison Sites Click to view larger Figure 9.1 Prices and Profits as Function of Participation Rate of Seller j. The expression for the profits of firm j is similar. Taking the first-order conditions with respect to pi and pj yields a system of equations that characterizes the equilibrium prices. Unfortunately it is difficult to derive a closed-form expression for these prices. Figure 9.1(a), however, shows for selected parameter values that the equilibrium price of firm i decreases in α j.10 Intuitively, as α jincreases, the (p. 243) probability a firm meets a competitor at the comparison site goes up and this fosters competition. As shown in Figure 9.1(b) the profits of firm i also decrease in α j. We now study the behavior of consumers in the market. If a user does not register with the comparison site, she just visits a firm at random. Therefore, the expected utility from this strategy is (3) where γ is Euler's constant. The utility a consumer obtains when remaining unregistered with the comparison site should increase in α i and α j because of the increased competition in the comparison site. If a user registers with the intermediary, her expected utility is (4) Armed with these equations, we can study the continuation game equilibria. Basically, there may be two kinds of equilibrium. An equilibrium with full consumer participation (μ* = 1) (and either full or partial firm participation), or an equilibrium with partial consumer participation (μ* 〈 1) (and either full or partial firm participation). In both types of equilibria, if the firms mix between advertising and not advertising the advertising probabilities must solve In the second type of equilibrium, if the users mix between participating and not participating, the participation probability μ* must be the solution to Page 14 of 21 Comparison Sites In the first stage of the game the owner of the comparison site chooses the pair of fees {a, s} such that her profits are maximized. The profit of the intermediary is given by (5) (p. 244) 3.2.1. Horizontal Product Differentiation It is convenient to first assume products are not vertically differentiated, i.e., δi = δj = δ. This means firms are symmetric and therefore we only need to consider two participation probabilities, α and μ. This case has been studied by Galeotti and Moraga-González (2009) for the N − firm case and a uniform distribution of match-parameters. Galeotti and Moraga-González (2009) show that in SPE the comparison site chooses firm and user fees so that firms and consumers join with probability one. That this also holds for a double exponentially distributed match-parameter can be seen in Table 9.7. The table shows that when either α or μ are less than one the profits of the intermediary can go up by increasing the fees to firms, consumers, or both. As a result, the intermediary will set both a and s such that all agents participate, that is, α = μ = 1. The intuition is similar to that in the model with homogeneous products. Suppose users mix between registering with the comparison site and not doing so. A raise in their registration fees makes them less prone to participate. To restore equilibrium firms should be more active in the comparison site. A higher participation rate of the firms can then be consistent with the expectation that consumers also participate more often. These cross-group externalities imply the increasing shape of the profits function of the comparison site in s. Therefore, the intermediary will continue to increase s till either α = 1 or μ = 1. When α = 1, the table shows how an increase in the firms fee increases consumer participation, lowers the price as well as firm participation. Profits of the comparison site increase anyway because consumer demand for participation is more elastic than firm demand for participation. Since an increase in a decreases firm participation, this relaxes the μ = 1 constraint and then the intermediary can increase again the consumer fee. This process continues until the intermediary extracts all the rents in the market up to the value of the outside option of the agents. We note that this insight is (p. 245) not altered if the retailers are allowed to price discriminate across marketplaces (see Galeotti and Moraga-González, 2009). Table 9.7 Model with Horizontal Product Differentiation a s p* α μ πi u* Π 0.050 0.100 1.552 0.681 0.108 0.253 1.030 0.079 0.050 0.150 1.550 0.752 0.110 0.252 1.031 0.092 0.050 0.200 1.547 0.825 0.113 0.251 1.032 0.105 0.050 0.313 1.541 1.000 0.121 0.249 1.036 0.138 0.100 0.313 1.518 0.993 0.242 0.215 1.044 0.274 0.200 0.313 1.477 0.982 0.487 0.146 1.059 0.545 0.400 0.313 1.412 0.963 0.980 0.006 1.083 1.078 0.400 0.300 1.405 0.989 0.992 0.002 1.088 1.119 0.401 0.337 1.401 1.000 1.000 0.000 1.090 1.139 Notes: The quality parameter δ is set to 1. One important assumption behind the efficiency result is that agents are all ex-ante symmetric. If for example firms are exante heterogeneous, a monopolist intermediary may induce a suboptimal entry of agents into the platform (see Nocke, Peitz and Stahl, 2007). Click-through fees If we allow for a click-through fee c in addition to an implicit and exogenous fixed cost k of advertising at the comparison site, Page 15 of 21 Comparison Sites the expected profits of a firm i are The first part of this profits expression comes from the consumers who participate in the comparison site. These consumers click on firm i's product to get directed to firm 's website when firm i is the only advertising firm and consumers prefer product i over the outside option, or when firm j also advertises its product but consumers prefer product i over product j and the outside option. The consumers who do not register with the platform buy at random and therefore do not generate any costs of clicks. The formula for the profits of the rival firm is similar. The utility consumers obtain when they opt out of the platform is the same as equation (3), while the utility they get when they register with the intermediary is the same as in equation (4). As we have done above, if firms mix between advertising at the platform and not advertising, it must be that and . If consumers mix between registering with the intermediary and opting out it must be that u*(1) * u*(0). The platform's profits stem from consumer subscriptions and click-through traffic. Therefore, The following table shows that the behavior of the profits function of the intermediary with click-through fees is qualitatively similar to the one with fixed fees. Suppose firms and consumers opt out with strictly positive probability. Then, for a (p. 246) given click-through fee, the intermediary can increase the consumer subscription fee and obtain a greater profit. Likewise, for a given subscription fee, the intermediary can increase the consumer subscription fee and increase benefits. As a result, firms and consumers must all participate surely. Table 9.8 Model with Horizontally Differentiated Products and CPC k c s p* α μ π* u* Π 0.200 0.500 0.050 1.763 0.655 0.595 0.115 0.983 0.292 0.200 0.500 0.100 1.759 0.727 0.608 0.111 0.979 0.342 0.200 0.500 0.150 1.755 0.801 0.622 0.107 0.975 0.392 0.200 0.600 0.150 1.824 0.817 0.668 0.094 0.957 0.488 0.200 0.800 0.150 1.985 0.855 0.771 0.065 0.914 0.720 0.200 1.000 0.175 2.179 0.940 0.891 0.031 0.856 1.044 0.200 1.186 0.182 2.386 1.000 1.000 0.000 0.800 1.369 Notes: The quality parameter δ is set to 1. Page 16 of 21 Comparison Sites 3.2.2. Vertical Product Differentiation We now consider the case in which firms sell products that are also vertically differentiated. Without loss of generality we normalize firm j's quality level to zero and assume firm i offers higher quality than firm j, that is, δi 〉 δj = 0. We first hypothesize that at the equilibrium fees a and s, consumers participate with probability less than one, so μ 〈 1. Moreover, for large enough differences in quality, the high-quality firm i participates with probability one and firm j participates with probability less than one, α i = 1 so and α j ∈ (0,1). If this is so it must be the case that Since we are assuming the intermediary has to charge the same fee a to both firms, such an equilibrium could arise if the intermediary sets her fees such that the low-quality firm is indifferent between participating or not participating—for large enough differences in quality this implies that the high-quality firm will always obtain higher profits by advertising on the comparison site than by selling to the non-participating consumers. From these equations we can obtain the equilibrium participation probabilities (α j,μ). Table 9.9 shows the behavior of firm and user participation rates and the profits of the intermediary when the quality difference Δ = δi− δj = 1. Starting from a (p. 247) relatively low level of a and s, an increase in s increases both consumer and firm participation. As a result, the intermediary will continue to increase s till α j is one. From that point onward, an increase in the firms’ fee increases the user participation rate and lowers the participation of firm j so the intermediary again finds it profitable to raise the consumer fees. This process continues until firm j and the consumers all participate with probability one. The intermediary maximizes its profit at a = 0.188 and s = 0.281. Table 9.9 Model with Vertical Product Differentiation (Δ = 1) a s pi pj αi αj μ πi πj u Π W 0.100 0.150 1.556 1.242 1.000 0.281 0.520 0.454 0.134 0.928 0.206 1.722 0.100 0.200 1.544 1.241 1.000 0.573 0.522 0.441 0.133 0.931 0.262 1.767 0.100 0.250 1.531 1.241 1.000 0.868 0.524 0.427 0.132 0.934 0.318 1.813 0.100 0.272 1.525 1.241 1.000 1.000 0.525 0.421 0.132 0.937 0.343 1.833 0.150 0.272 1.501 1.214 1.000 0.974 0.792 0.348 0.058 0.944 0.512 1.862 0.188 0.272 1.481 1.188 1.000 0.952 0.988 0.293 0.000 0.950 0.639 1.882 0.188 0.281 1.476 1.188 1.000 1.000 1.000 0.288 0.000 0.952 0.657 1.897 0.567 −0.122 1.567 1.193 1.000 0.000 1.000 0.000 0.000 1.148 0.445 1.594 Notes: The quality parameter δi is set to 1, while δj = 0. So far we have looked at an equilibrium in which the fees are set such that both firms will participate. However, the intermediary could also set its fees such that only the high-quality firm participates, which means it must be the case that The last line of Table 9.9 shows that if the intermediary sets her fees such that only the high-quality firm participates, profits Page 17 of 21 Comparison Sites are lower than if it lets the two firms participate. Nevertheless, for relatively large differences in quality it might be the case that the intermediary prefers only the highly quality firm to advertise on its website. For instance, when Δ = 3 the comparison site maximizes profits by setting a and s in such a way that firm j does not find it profitable to advertise on the platform, while firm i participates with probability one. This can be seen in Table 9.10 where we describe the behavior of firm and consumer decisions and the implications for the comparison site's profits forΔ = 3. As shown in the table, since the high-quality firm is the only firm active at the comparison site, for all a and s the price found on the intermediary will be the same. Starting from relatively low (p. 248) fee levels a higher s leaves the participation rate of the consumers unchanged—α i is changed in such way to keep consumers indifferent between participating and not participating. Increasing a leads to higher participation of the consumers, and as such to higher profits for the intermediary. As shown in the table the intermediary maximizes her profits when a = 1.557 and s = 0.347 (in bold). Comparing the last two lines of this table shows that in this case setting a relatively low a such that both firms will participate will generate lower overall profits for the platform in comparison to setting a relatively high a such that only the high-quality firm participates. However, as shown by the last column of Table 9.10 welfare W is higher when everyone participates. Table 9.10 Model with Vertical Product Differentiation (Δ = 3) a s pi pj αi αj μ πi πj u Π W 1.000 0.100 2.557 1.234 0.737 0.000 0.642 0.557 0.100 1.169 0.802 2.628 1.000 0.200 2.557 1.225 0.844 0.000 0.642 0.557 0.100 1.169 0.972 2.799 1.000 0.300 2.557 1.215 0.950 0.000 0.642 0.557 0.100 1.169 1.143 2.969 1.000 0.347 2.557 1.210 1.000 0.000 0.642 0.557 0.100 1.169 1.223 3.050 1.300 0.347 2.557 1.172 1.000 0.000 0.835 0.257 0.046 1.169 1.590 3.063 1.500 0.347 2.557 1.138 1.000 0.000 0.963 0.057 0.010 1.169 1.835 3.072 1.557 0.347 2.557 1.127 1.000 0.000 1.000 0.000 0.000 1.169 1.905 3.074 0.115 0.490 2.388 1.115 1.000 1.000 1.000 1.273 0.000 1.241 0.720 3.235 Notes: The quality parameter δi is set to 3, while δj = 0. 4. Concluding Remarks and Open Research Lines The emergence of Internet shopbots and their implications for price competition, product differentiation, and market efficiency have been the focus of this chapter. We have asked a number of questions. How can a comparison site create value for consumers? Do firms have incentives to be listed in comparison sites? Under which conditions are prices listed in these sites dispersed? Do comparison sites enhance social welfare? Can comparison sites generate efficient allocations? To answer these questions we have developed a simple model of a comparison site. The intermediary platform tries to attract (possibly vertically and horizontally differentiated) online retailers and consumers. The analysis of the model has revealed that product differentiation and price discrimination play a critical (p. 249) role. For the case of homogeneous product sellers (Baye and Morgan, 2001), if the online retailers cannot charge different on- and off-the-comparison-site prices, then the comparison site has incentives to charge fees so high that some firms are excluded. The fact that some firms are excluded generates price dispersion, which creates value for consumers. This value, in turn, can be extracted by the comparison site via consumer participation fees. The market allocation is not efficient since products are sold at prices that exceed marginal cost. By contrast, when on- and off-the-comparison-site prices can be different, the comparison site has an incentive to lure all the players to the site, which generates an allocation that is efficient. When online retailers sell products that are horizontally differentiated, the comparison site creates value for consumers by aggregating product information. In this way, the comparison site easily becomes a more attractive marketplace than the decentralized one (Galeotti and Moraga-González, 2009). In equilibrium, even if firms cannot price discriminate the comparison site attracts all the players to the platform and an efficient outcome ensues. Page 18 of 21 Comparison Sites Allowing for vertical product differentiation brings interesting additional insights. The platform faces the following trade-off. On the one hand, it can attract high-quality and low-quality producers to the platform so as to increase competition, aggregate information and generate rents for consumers that are ultimately extracted via registration fees. Alternatively, the comparison site can set advertising fees that are so high that low-quality sellers are excluded, thereby creating value for the top sellers. When quality differences are large, the latter strategy pays off. The comparison site excludes low quality from the platform, and grants higher rents for advertising sellers. Part of these rents are ultimately extracted by the comparison site. Some value for consumers is destroyed and the resulting allocation is inefficient. Along the way, we have kept things simple and therefore left aside important issues. One important assumption has been that firms and consumers could only use a single platform as an alternative to the search market. In practise, multiple comparison sites exist and questions about how they compete with one another and their sustainability in the long-run arise. Moreover, if retailers are ex-ante differentiated, one aspect worth to investigate is whether they distribute themselves across platforms randomly of else they are sorted in a top-down way across them. One practice we observe in recent days is that retailers are given the possibility to obtain priority positioning in comparison sites’ lists after paying an extra fee. Adding this choice variable to the problem of the platform's manager would probably complicate matters much. One way to address this issue is to use the framework put forward in the literature on position auctions. Priority positions could be auctioned as it is the case in search engines advertising. For instance, Xu, Chen and Whinston (2011) study a model where firms sell homogenous products and bid for prominent list positions. They relate the extent of market competitiveness to willingness to bid for prominent positions. Another realistic feature we have ignored throughout is incomplete information. When quality is private information of the retailers, the (p. 250) problem of the platform is how to screen out good from bad thereby creating value for consumers. For a mechanism design approach in the context of search engines to this problem see Gomes (2011). Finally, another simplifying assumption has been that search costs are negligible within a platform. When searching the products displayed in the price comparison site is still costly for consumers, platforms may have incentives to manipulate the way the alternatives are presented, thereby inducing more search and higher click turnover (see Hagiu and Jullien, 2011). On the empirical side, much research remains to be done. In fact, not much is yet known about the extent to which theoretical predictions are in line with the data, especially in settings where product differentiation is important. Comparison sites can potentially provide a wealth of data. Detailed data from comparison sites may be helpful to learn how exactly consumers search for products. Some consumers might sort by price, while others may sort by quality ratings or availability. This type of information may give firms indications about the best way to to present their product lists. Click-through data can facilitate the estimation of structural models of demand. Finally, using data from bids for prominent positions may be a useful way to estimate the characteristics of the supply side of the market. Acknowledgments We thank Martin Peitz, Joel Waldfogel, and Roger Cheng for their useful comments. Financial support from Marie Curie Excellence Grant MEXT-CT-2006–042471 is gratefully acknowledged. References An, Y., M.R. Baye, Y. Hu, J. Morgan, and M. Shum (2010), “Horizontal Mergers of Online Firms: Structural Estimation and Competitive Effects,” mimeo. Anderson, S.P. (2012), “Advertising and the Internet,” In: The Oxford Handbook of the Digital Economy, M. Peitz and J. Waldfoegel (eds.). New York/Oxford: Oxford University Press. Anderson, S.P. and S. Coate (2005), “Market Provision of Broadcasting: A Welfare Analysis,” Review of Economic Studies, 72, 947–972. Armstrong, M. (2006), “Competition in two-sided markets,” Rand Journal of Economics, 37, 668–691. Armstrong, M. and J. Wright (2007), “Two-Sided Markets, Competitive Bottlenecks and Exclusive Contracts,” Economic Theory 32, 353–380. Athey, S. and G. Ellison (2011), “Position Auctions with Consumer Search,” Quarterly Journal of Economics 126, 1213–1270. Baye, M.R. and J. Morgan (2001), “Information Gatekeepers on the Internet and the Competitiveness of Homogeneous Product Markets,” American Economic Review 91, 454–474. Baye, M.R., J. Morgan, and P. Scholten (2004), “Price Dispersion in the Small and in the Large: Evidence from an Internet Price Comparison Site,” The Journal of Industrial Economics 52, 463–496. Page 19 of 21 Comparison Sites Baye, M.R., J. Morgan, and P. Scholten (2006), “Information, Search, and Price Dispersion,” Chapter 6 in Handbook in Economics and Information Systems, Volume 1 (T. Hendershott, Ed.), Amsterdam: Elsevier. (p. 252) Baye, M.R., R. Gatti, P. Kattuman, and J. Morgan (2006), “Did the Euro Foster Online Price Competition? Evidence from an International Price Comparison Site,” Economic Inquiry 44, 265–279. Baye, M.R. and J. Morgan (2009), “Brand and Price Advertising in Online Markets,” Management Science 55, 1139–1151. Belleflamme, P. and E. Toulemonde (2009), “Negative Intra-Group Externalities in Two-Sided Markets,” International Economic Review, 50–1, 245–272. Brown, J. and A. Goolsbee (2002), “Does the Internet Make Markets More Competitive? Evidence from the Life Insurance Industry,” Journal of Political Economy 110, 481–507. Brynjolfsson, E. and M.D. Smith (2000), “Frictionless Commerce? A Comparison of Internet and Conventional Retailers,” Management Science 46, 563–585. Caillaud, B. and B. Jullien (2003), “Chicken and Egg: competition among intermediation service providers,” Rand Journal of Economics 34, 309–328. Chen, Y. and C. He (2011), “Paid Placement: Advertising and Search on the Internet,” Economic Journal 121, F309–F328. Demsetz, H. (1968), “The Cost of Transacting,” Quarterly Journal of Economics, 33–53. Van Eijkel (2011), “Oligopolistic Competition, OTC Markets and Centralized Exchanges,” unpublished manuscript. Ellison, G. and S. Fisher Ellison (2009), “Search, Obfuscation, and Price Elasticities on the Internet,” Econometrica 77, 427– 452. Evans, D.S. (2003), “The Antitrust Economics of Two-Sided Markets,” Yale Journal on Regulation 20, 325–382. Galeotti, A. and J.L. Moraga-González (2009), “Platform Intermediation in a Market for Differentiated Products,” European Economic Review, 54, 417–28. Gehrig, T. (1993), “Intermediation in Search Markets,” Journal of Economics and Management Strategy 2, 97–120. Gomes, R. (2011), “Optimal Auction Design in Two-Sided Markets,” unpublished manuscript. Hagiu, A. and B. Jullien (2011), “Why Do Intermediaries Divert Search?” Rand Journal of Economics 42, 337–362. Koulayev, S. (2010), “Estimating Demand in Online Search Markets, with Application to Hotel Bookings,” unpublished manuscript. Nocke, V., M. Peitz and K. Stahl (2007), “Platform Ownership,” Journal of the European Economic Association 5, 1130–1160. Rochet, J-C. and J. Tirole (2003), “Platform Competition in Two-Sided Markets,” Journal of the European Economic Association 1, 990–1029. Rochet, J.-C., and J. Tirole (2006), “Two-Sided Markets: A Progress Report,” Rand Journal of Economics 37, 645–67. Rubinstein, A. and A. Wolinsky (1987), “Middlemen,” The Quarterly Journal of Economics 102, 581–94. Smith, M.D. and Brynjolfsson E. (2001), “Customer Decision-Making at an Internet Shopbot: Brand Matters,” Journal of Industrial Economics 49, 541–558. Spiegler, R. (2011), Bounded Rationality and Industrial Organization, Oxford University Press. Spiegler, R. and K. Eliaz (2011), “A Simple Model of Search Engine Pricing,” Economic Journal 121, F329–F339. Spulber, D. (1999), Market Microstructure: Intermediaries and the Theory of the Firm, Cambridge University Press. (p. 253) Weyl, E.G. (2010), “A Price Theory of Multi-Sided Platforms,” American Economic Review 100, 1642–1672. Watanabe, M. (2010), “A Model of Merchants,” Journal of Economic Theory 145, 1865–1889. Yavas, A. (1994), “Middlemen in Bilateral Search Markets,” Journal of Labor Economics 12, 406–429. Yavas, A. (1996), “Search and Trading in Intermediated Markets,” Journal of Economics and Management Strategy 5, 195– Page 20 of 21 Comparison Sites 216. Notes: (1.) Weyl (2010) presents a general model. The literature on two-sided markets has typically focused on situations where network effects are mainly across sides. As we will see later, for profit-making of comparison sites, managing the externalities within retailers is of paramount importance. (2.) Alternatively, and without affecting the results qualitatively, these consumers can be seen as buying from a “local” firm, or from a firm to which they are loyal. (3.) Later in the chapter, we show that our main results hold when per-click fees are used. As mentioned in the Introduction, some comparison sites may obtain the bulk of their revenue from the selling of advertising space instead. This alternative business model is considered elsewhere in this volume. (4.) In this chapter we will allow for price discrimination across, but not within, marketplaces. That is, the firms may be able to post on the price comparison site a price different from the one they charge to local/loyal consumers. Things are different when retailers can price discriminate across loyal consumers. This is often the case in over-the-counter (OTC) markets, where the bargaining institution is predominant. For a model that studies the role of price discrimination over-the-counter in a two-sided market setting see Van Eijkel (2011). (5.) Matching values are realized only after consumers visit the platform or the individual shops. This implies that ex-ante all consumers are identical. This modelling seems to be appropriate when firms sell newly introduced products. (6.) It is well known that in models with network externalities multiple equilibria can exist. In our model there is always an equilibrium where no agent uses the comparison site. This equilibrium is uninteresting and will therefore be ignored. (7.) To be sure, the model of Baye and Morgan (2001) is more general because they study the N-retailers case and because consumers have elastic demands. In addition, for later use, we have assumed that consumers who register with the platform cannot visit the local shop any longer. The results do not depend on this assumption and we will also make it here for overall consistency of our chapter. (8.) For the moment, we are ignoring the possibility firms advertise on the platform a price different from the one they charge to the non-participating consumers. (9.) One reason why CPC tariffs may recently have become more widely used is that they involve less risk for the platform. An, Baye, Hu, Morgan, and Shum (2010) also allow for a click-trough fee in a Baye and Morgan (2001) type model, but do not model the optimal fee structure of the comparison site. (10.) The parameter values used in Figure 9.1 are μi = μj = 1, α i = 1, δi = 1, and δi = 0. Jose-Luis Moraga-Gonzalez Jose-Luis Moraga-Gonzalez is Professor of Microeconomics at VU University-Amsterdam and Professor of Industrial Organization at the University of Groningen. Matthijs R. Wildenbeest Matthijs R. Wildenbeest is Assistant Professor of Business Economics and Public Policy at the Kelley School of Business, Indiana University. Page 21 of 21 Price Discrimination in the Digital Economy Oxford Handbooks Online Price Discrimination in the Digital Economy Drew Fudenberg and J. Miguel Villas-Boas The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0010 Abstract and Keywords This article addresses the effects of price discrimination that is based on more detailed customer information, both under monopoly and under competition. The activities of the firms, namely using information gained from consumers through their purchase decisions, raise the possibility that consumers may understand that the decisions which they take may impact the options that they will have available in the future periods. It is noted that competing firms can potentially gain more from learning the valuation of consumers than not learning, which could be a force for greater competition for consumers. In some markets, firms learn consumer characteristics that directly influence the cost of servicing them. For situations with competition, if the competitors are aware that firms have purchase history information, more information may actually result in more intense competition after the information is gained. Keywords: price discrimination, customer information, monopoly, competition, competing firms, consumers 1. Introduction With the developments in information technology firms have more detailed digital information about their prospective and previous customers, which provides new mechanisms for price discrimination. In particular, when firms have information about consumers’ previous buying or search behavior, they may be able to use this information to charge different prices to consumers with different purchase histories. This form of price discrimination is present in several markets, and it may become increasingly important with greater digitalization of the market transactions. The development of the information technologies and web-browser cookies allows firms to collect, keep, and process more information about consumers, and can affect the prices and products offered by firms to different consumers. This article discusses the effects of price discrimination based on information gained by sellers from consumer purchase histories under monopoly and competition. When price discrimination is based solely on purchase histories, it is called behavior-based price discrimination. This article also discusses the effect of finer (p. 255) information at the customer level on price discrimination practices. Having more information about the consumers’ product valuation (through purchase histories or other methods) helps the firm extract more surplus from the consumers. However, consumers may anticipate this possibility, and therefore alter their initial purchases. Firms may also be proactive in sorting consumers that purchase in the early periods in order to gain more information on their preferences. With competition, more information about consumers may lead to pricing as if there were less product differentiation as firms target each separate price to a less heterogeneous consumer pool. This effect can lead to lower equilibrium prices. Nevertheless, competing firms can potentially benefit from customer recognition in spite of lower equilibrium prices if the information leads to changes in the products sold—either higher sales of same-cost products or a switch of sales to lower cost products—or to sales to consumers who are less costly to Page 1 of 15 Price Discrimination in the Digital Economy serve. This article follows closely the results by Hart and Tirole (1988), Schmidt (1993), and Villas-Boas (2004) on monopoly, and the results by Villas-Boas (1999) and Fudenberg and Tirole (2000) for competition, while discussing the effects of switching costs considered in Chen (1997) and Taylor (2003). For a more extended survey of the behavior-based price discrimination literature see Fudenberg and Villas-Boas (2006).1 The article is organized in the following way. The next section considers the potential short-term market effects of firms having more information about their consumers. We consider both the case of monopoly and of competition, and for each case we consider the possibility of no information being tracked, the possibility of information being solely based on purchase histories (with either fixed consumer preferences, changing consumer preferences, or entry of new generations of consumers), and the possibility of firms learning the preferences of their customers through other channels in addition to purchase choices. Section 3 looks at the strategic behavior of consumers when they foresee that their purchase decisions affect the market opportunities available to them in the future, because of the information gained by firms. Section 4 discusses the strategic actions of firms when gaining information about consumers through their purchases. Section 5 discusses several extensions, and Section 6 concludes. 2. Information About Consumers Consider a firm selling a product to a set of heterogeneous consumers of mass one. Suppose consumers have a valuation for the product v with cumulative distribution function F(v), density f(v), with the property that v[1 − F(v)] is quasi-concave in v. Suppose marginal costs are zero, and consider the firm being able to know something about the valuation of some consumers, and being able to charge consumers differently depending on its information. Suppose also that there are no (p. 256) strategic issues about the firm learning more from consumers in this purchase occasion, and that consumers are not concerned about any effects on their future payoffs of their decisions to buy or not to buy at this purchase occasion. This can be seen as the ex-post situation after the firm learned something about the consumers in the previous periods. This can also be seen as the situation in the second period of a two-period model, where the firm learned something from the consumers that bought in the first period. There are several modeling possibilities of what the firm knows about the valuation ν of each consumer. One possibility is that the firm does not know anything about the valuation of each consumer and just knows that the cumulative distribution of ν is F(ν). This could be the case when the firm is not able to keep track of which consumers bought the product in the previous period or there are no information technologies to keep track of the consumer purchase histories. This could also be the case if the valuations of each consumer were independent over time. In this “no information” case, as the firm does not know anything about ν per consumer, the firm just charges a price P which is the same for all consumers. Since consumers are not concerned about the firm learning their valuation while purchasing the product, they buy the product as long as ν ≥ P. The profit-maximizing price P is then the optimal static monopoly price P* = argmaxp P[1 − F(p)]. Another possibility is that consumers have the same valuation from period to period, and entered the market in the previous period, and the firm is able to identify which consumers bought the product in the previous period (with consumers being around for two consecutive periods).2 Suppose that there was a demand x in the previous period and that the consumers who bought the product in the previous period are the ones with the highest valuation. Then, we know that all consumers who bought in the previous period have valuation greater or equal to some valuation v* which is determined by x = 1 − F(v*). The firm can then charge two prices, one price for the consumers who bought in the previous period, which the firm knows to have valuations above v*, and another price for the consumers that did not buy in the previous period, which the firm knows to have valuation less than v*. Let us denote the price to the consumers who bought in the previous period as Po, and the price to the new consumers as Pn. For the price to the previous customers, if v* ≥ P* then the firm just charges Po = v*, as the optimal price for those consumers without any constraints on their valuations would have been P*. For the same reason, if v* 〈 P* then the firm chooses to charge Po = v*. One can then write Po = max[v*, P*] which considers both cases. * Page 2 of 15 Price Discrimination in the Digital Economy For the price to the new consumers the firm chooses to charge pn = argmaxp P[F(v*) − F(p)], which accounts for the fact that the firm is only targeting the consumers with valuation below v*. Considering both the price to the previous customers and the price to the new consumers, the firm is better off in this period with the ability to identify the consumers that bought in the previous period, as it is able to better extract the consumer surplus through price discrimination. This is a continuous setup of the two-type model considered in Hart and Tirole (1988) and Villas-Boas (2004). (p. 257) One variation of this possibility is that consumers’ valuations change through time. When this happens the firm may potentially want to charge a lower price to the consumers who purchased in the previous period, and a higher price to the new customers. One example of changing preferences is one where with some probability ρ consumer change preferences from one period to the next, and in the next period the consumer's valuation is an independent draw from the distribution F(v). Consider the effect of this on the profit-maximizing price for the previous customers. If v* ≥ P*, the firm's profit from the previous customers for a price Po ≤ v* is Po[(1 − ρ)1 − F(v*)] + ρ[1 − F(v*)]1 − F(Po)]}. One can obtain that for sufficiently small ρ we have the optimal Po = v*. After a certain threshold value of ρ, the optimal price for the previous customers, Po, starts decreasing from v* with increases in ρ, and keeps on decreasing until it reaches P* when ρ = 1. If v* 〈 P*, the optimal price for the previous customers is always P*, independent of ρ. Regarding the price for the new customers, the firm's profit from the new customers is Pn{(1 − ρ)[F(v*) − F(Pn)] + ρF(v*)[1 − F(Pn)]}. One can then obtain that the optimal Pn is increasing in ρ, reaching P* when ρ = 1. Note that a firm is still strictly better off with this possibility of identifying its past consumers as long as ρ 〈 1. Another interesting variation is the case when a new generation of consumers comes into the market, so that the firm knows that some of their new consumers may have a high valuation for the product, and may be in the market for future periods. This case is considered in Villas-Boas (2004) with overlapping generations of consumers. In comparison to the model above, this variation adds a force for the firm to raise the price charged to the new consumers, as some of the new potential consumers have a high valuation for the firm's product. Finally, another possibility is that the firm is able to learn something about the valuations of the consumers that bought from it, and charge them differently because of their valuation. Note that this case is not “pure” behaviorbased price discrimination as in the case above. In the case above, the firm only learns whether a consumer valuation is such that a consumer makes a certain choice (either buy or not buy). In some cases a firm may be able to learn something more about a consumer's valuation from a consumer that bought the product. For example, a consumer who chose to buy the product may reveal to the firm during the purchase process some information about her/his preferences and valuation, perhaps during the exchange of information with a salesperson, or the interaction of the consumer with a website through the search process, or after the purchase during the servicing of the consumer. In addition to learning about a consumer's valuation, the firm may learn about the cost of servicing the consumer. This can be particularly important in insurance or credit markets, where a firm may learn after purchase whether a customer is more or less likely to have an accident, or more or less likely to default on its debt. It can also be important in markets where customers can vary in the cost of servicing their purchases. For example, a cell telephone company knows how often each of its customers calls their service center while it does not have that information for noncustomers. In these cases a firm can benefit even more in this period from having information on the consumers who bought in the previous period. (p. 258) An extreme version of learning about the previous customers’ valuations is when the firm learns all the valuation of the consumers that bought the product, and can perfectly price discriminate between them. In this case, previous consumers with valuation v ≥ v* would each be charged a price Po(v) = v, each ending up with zero surplus in this period. For consumers with valuation v 〈 v*, the ones that did not purchase in the previous period, the firm does not know the consumers’ valuation, and therefore the price that they are charged is the same as obtained above, Pn argmaxp P[F(v*) − F(p)]. In this case, consumers end up with lower surplus after revealing their preference information, and the seller is better off, than in the case where the seller has only information about the consumer purchase histories. Competition Page 3 of 15 Price Discrimination in the Digital Economy Consider now a variation of the model where there is competition, as in Fudenberg and Tirole (2000) and VillasBoas (1999). Suppose that we have a duopoly with Firms A and B competing in a market. A consumer with valuation v for Firm A has valuation z − v for Firm B, where z is a constant, and such that 2v represents the relative preference of a consumer for good A over good B. Denoting the support of v as [v̱, vÌ„], the two firms are symmetric if z = vÌ„ + v̱, and f(v) is symmetric around , which we assume. Suppose also that v̱ is high enough such that the full market is covered. Then, as above, suppose conditions under which (1) firms do not know anything about the consumers’ valuations, (2) firms can identify the consumers that bought from them, or (3) consumers learn the valuations of the consumers that bought from them. If firms do not know anything about consumers valuations they will just set one price for all consumers. For Firm A, charging price PA, with the competitor charging the price PB, the demand is obtain similarly the demand for Firm B, and obtain that in equilibrium, Fv is the uniform distribution this reduces to . We can . For the case where . If firms are able to identify the consumers that bought from them in the previous period, consumer preferences remain unchanged, and there are no new consumers in the market, a firm may know that the consumers that bought from it have a relative preference for its product, and may then be able to charge one price to its previous customers, and another price to the previous customers of the competitor. Given the demand that a firm obtained in the previous period, it can identify that the threshold valuation of a consumer that bought from it in the previous periods. Denote this threshold valuation as v*, the valuation for Firm A, such all consumers with valuation v 〉 v* for Firm A chose Firm A, and all consumers with (p. 259) valuation v 〉 v* for Firm A chose Firm B (as, by definition, their valuation is greater than z − v* for Firm B). We can then separate the problem in finding the equilibrium within the previous customers for Firm A, and within the previous customers for Firm B. For the previous customers of Firm A, Firm A can charge price PoA and Firm B can charge price PnB. The price for Firm A of maximizing the profits from its previous customers is determined by and the one for Firm B is determined by . Solving these two equations together one can obtain the equilibrium prices. For example, for the uniform distribution example one can obtain and . Under general conditions (for example, satisfied by the uniform distribution) one can obtain that in equilibrium PnB is smaller than PoA, and that PoA is smaller than the equilibrium price under no information, . The intuition for the first result is that Firm B has to price lower because it is trying to attract consumers that have a higher preference for Firm A. The intuition for the second result is that Firm A's price has to respond to a lower price from Firm B, and given strategic complementarity in prices, Firm A's price is lower than the no information equilibrium prices. A similar analysis can be obtained for Firm B's previous customers. As the market is fully covered, we can then obtain that all consumers pay a lower price than in the no-information case, and that industry profits in the current period (second period of a two-period model) are now lower. That is, after learning the information about consumers, competition leads to lower profits, while the result was different under monopoly. Page 4 of 15 Price Discrimination in the Digital Economy This is for example the result in Fudenberg and Tirole (2000). This result is similar to the one in competition with third-degree price discrimination, where there is a market force leading to lower profits than in competition with uniform pricing, because there is now less differentiation across firms for the consumers being served for each pair of prices. An example of this is Thisse and Vives (1988), where firms know the exact location of the consumers and compete with location-based price discrimination. Note also that this result is for the industry profits after the information is revealed. It may still be possible in this setting that the present value of industry profits is greater when considered before information is obtained from consumers, if consumers foresee that their choices in the earlier periods affect the offers that they will get from the firms in future periods. This effect is studied in Section 4. One particular characteristic of the framework above is that firms know as much about their previous customers as about the previous customers of the competitors. That is, if a new consumer approaches a firm the firm knows that the consumer was a customer of the competitor in the previous period. One way to take out this characteristic from the model is to allow new generations of consumers to come into the market as considered in Villas-Boas (1999). In that case, a firm knows that a pool (p. 260) of new consumers is composed of old consumers that do not have a high valuation for its product (and a high valuation for the competitor), and by a new generation of consumers, where several of them may have a high valuation for the firm's product. This is then a force for the firm to charge a higher price to its new customers (as some of them may have a high valuation), which then, by strategic complementarity of prices, leads the competitor's price to its previous customers to be higher. Another variation, as considered above for the monopoly case, is that a fraction of consumers change preferences from the previous period to the new period (e.g., Chen and Pearcy, 2010). That is, some of a firm's previous customers may now have a lower valuation for the firm's product than in the previous period, and some of the competitor's previous customers may now have a higher valuation for the firm's product. This will then lead the firm, in comparison to the case when the preferences do not change, to lower the price for its previous customers, and raise the price to its new customers. Another possibility for a firm to learn about its previous customers’ valuations, also as discussed above for the monopoly case, is to learn something about the valuation of each consumer during the purchase process or when servicing the consumer post-purchase, for the consumers that bought from the firm. An extreme version of this case is when the firm learns perfectly the valuation of its previous customers. For Firm A, the price for a previous customer with a valuation v would then be the price (if greater than its marginal costs) that it would lead that consumer to be just indifferent between purchasing the product from Firm A at that price, and purchasing the product from Firm B at its price for the ne customers. That is, v = PoA(v) = z − v − PnB. Note that if , the competitor would find it profitable to attract some of the previous customers of Firm A, as Firm A does not charge a price below its marginal cost. If we have , then the equilibrium involves PnB = 0 (as marginal costs are assumed equal to zero) and PoA(v) = 2v − z. For the uniform distribution case one can still obtain that in this case profits are lower than in the no information case. One can however consider variations of this model where the firms’ learn the profitability of their previous customers (e.g., insurance markets, credit markets), which can potentially lead to more information leading to ex-post greater profits, as the informed firm may be able to extract more surplus from the different consumers (see, for example, Shin and Sudhir, 2010, for a study along these lines). 3. Strategic Behavior of Consumers The activities of the firms considered in the previous section, namely using information gained from consumers through their purchase decisions, are contingent on those decisions, which raises the possibility that consumers may understand (p. 261) that the decisions that they take may affect the options that they will have available in the future periods. In this section we assume that the firm does not have any information about any individual consumer prior to that consumer's decision, and further assume that the firm only uses this information for one more period after this consumer's decision (as in the analysis in the previous section). This could be seen as the consumer entering the market in this period and staying in the market for only two periods. The consumers Page 5 of 15 Price Discrimination in the Digital Economy discount the next period payoff with the discount factor δc ≤ 1. A consumer considers whether to buy the product now at a price pf and be offered a price po in the next period, or not to buy it this period, and be offered the product in the next period at the price pn. If a consumer with valuation v buys the product in this period and next period, he gets a surplus v − pf + δc (v − pn, 0). If the consumer decides not to buy this period, he will get a surplus δC max[ν − pi,0]. As noted in the preceding section, the marginal consumer who is indifferent between buying and not buying in the current period, v*, ends up with zero surplus in the next period, because it faces a price po ≥ v*. Then the valuation v* of a consumer indifferent between buying and not buying in the current period is determined by From this it is clear that if, for some reason, the price to the new consumers is expected to increase in the next period, pn ≥ pf , then all consumers with valuation greater than the current price charged, pf , should buy the product. Therefore, in this case we have v* = pf . This can occur, for example, when there is entry of new consumers in the next period, such that the firm chooses to raise its price to the new consumers. Villas-Boas (2004) shows that in such a setting the equilibrium involves price cycles, where in some periods the price to the new consumers is indeed higher than the price to new consumers in the previous period. If, on the other hand, the consumers expect the price to the new consumers to be lower next period, pn 〈 pf (which, as shown in the earlier section, happens when there is no entry of more consumers next period), then consumers will be strategic, and there are some consumers with valuation above the current price pf who prefer to wait for the next period, because they get a better deal next period. In this case the valuation of the marginal consumer indifferent between buying and not buying in the current period can be obtained as . Consumers with valuation v ∈ [pf , v*] would buy the product if they were myopic, but do not buy if they are strategic. This effect may hurt the firm, as fewer consumers buy in this period when the firm is able to practice behavior-based price discrimination, than when the firm does not keep information of the consumers that bought from it. This is known as the “ratchet effect” of the consumers loosing surplus (i.e., being charged a higher price) by revealing some information about their valuations (Freixas et al., 1985). (p. 262) If the firm is able to learn something about the valuation of each consumer during the purchase process, then the marginal consumer if buying in the current period also ends up getting zero surplus in the next period, and therefore, the valuation of the marginal consumer buying in the current period is obtained in the same way. In another variation of the model mentioned above, if preferences change through time, then a marginal consumer buying in the current may have a positive surplus in the next period if his preferences change to higher valuation than the price to the previous customers. Because some of the previous customers may lower their valuation the firm may also lower its price to the previous customers. These two effects are a force for more consumers to be willing to buy at a given price in the current period (lower price and consumers are less likely to have zero surplus in the next period of buying in the current period). In the extreme case where valuations are completely independent from period to period, the problem of the seller ends up being like a static problem in every period, and all consumers with valuation above the price charged in a current period buy the product in that period. Competition Under competition, if the firms are not able to identify their previous customers, then customers make choices in each period as if there are no future effects, and the market equilibrium is as characterized above for the no information case. If the firms are able to identify their previous customers, then in the competition model of the previous section, a marginal consumer has to decide between buying from Firm A and getting the competitors’ product in the next period at the price for its new customers, getting a surplus of v − pfA + δc (z − v − pnB), and buying from Firm B and getting product A in the next period at the price of its new customers, getting a surplus of z − v − pfB + δc (z − * Page 6 of 15 Price Discrimination in the Digital Economy v − pnA). The valuation v* for product A of the consumer indifferent between buying either product in the current period has then to satisfy making these two surpluses equal, leading to (2v* − z)(1 − δc ) = pfA − pfB + δc (pnB − pnA) where pnA and pnB are also a function of v*, given the analysis in the previous section. From above, if the marginal consumers decide to buy product A in the current period instead of product B, then they know that Firm B will charge a higher price to new consumers in the next period (as those new consumers have now a greater preference for Firm B). That means that a price cut by Firm A in the current period leads to fewer consumers switching to product A than if the next period prices were fixed, as consumers realize that by switching they will be charged a higher price in the next period from the product that they will buy, product B. That is, the possibility of behavior-based price-discrimination in the future makes the consumers less price sensitive in the current period. For the uniform distribution example we can obtain , where the (p. 263) demand for product A is composed of consumers with a valuation for product A of v ≥ v*. This lower price sensitivity may lead then to higher prices in the first period, as discussed in the next section. When some of the consumers’ preferences change from period to period the results above continue to hold, but we get closer to the no information case. When there is entry of new consumers, as described in the section above, the effect of current demand realization on future prices for new consumers is reduced. In that case, demand may become more price sensitive when firms can identify their previous customers. This is because the consumers at the margin know that they switch products from period to period. Then, their problem is the one in which order to get the products: product A in the current period followed by product B in the next period, or the reverse order. In the steady state of a symmetric equilibrium in an overlapping generations model, consumers become less concerned about this order as their discount factor δυ increases, so they become more and more price sensitive (Villas-Boas, 1999). This effect is also present in the case where there is no entry of new consumers, but in that case the effect mentioned above of greater prices in the next period for switching consumers dominates. When firms learn the consumers’ exact valuations from their previous customers, the marginal consumers in a symmetric equilibrium end up with a surplus equal to , as the poaching price is zero and a firm charges a price equal to 2v − z to its previous customers in the next period, as presented in Section 2. In this case, for the symmetric market, the marginal consumers who switch to product A know that they will be able to buy product B at a price that is below what they would get if they stayed with product B. This means that in this case the demand is more price sensitive when charging a lower price in the current period than in the case when firms did not learn about the valuations of their previous customers. For the uniform example, if pfA 〈 pfB, we have the marginal valuation for product A of a consumer that buys product A in the current period defined by . This shows for the uniform case that demand is more price sensitive than when firms do not learn about the valuations of their previous customers. 4. Firms Gaining Information About Consumers Now we consider the strategic decisions of the firms when their decisions change the information that firms gain in the current period. In order to focus exclusively on the strategic effects of gaining information, suppose that the firm does not currently know anything about the consumers’ valuations, and that it will only use the information obtained now for the next period, where payoffs will be discounted (p. 264) with discount factor δF. Note that if no information is gained by the consumers’ choices, there are no dynamic effects and the firm's decision is exactly as considered in Section 2 for the no information case. Monopoly We first examine the impact of various information structures under monopoly. Page 7 of 15 Price Discrimination in the Digital Economy Consider the case where the seller in the next period is able to recognize its previous customers so that it knows that their valuation for the product is above a certain threshold, and that the valuation of the other consumers is below that threshold. The problem of the seller (if the price to the new consumers falls in the next period, which can be shown to hold) is then where Pn(Pf ) is obtained as described in Section 2 as , and π2 (Pf ) is the profit in the next period, which, given the analysis in the preceding sections, can be presented as where the first term represents the profit next period from the consumers that purchased the product also in the current period, and the second term represents the profit next period from the consumers that did not purchase the product in the current period. For the case when firms and consumers discount the future in the same way, δc = δF = δ, one can show that at the optimum the valuation of the marginal consumers choosing to purchase in the current period, v*, is strictly greater than the valuation of the marginal consumers. As discussed above, this is because consumers are aware that prices may fall in the future, and therefore prefer to wait for the lower prices rather than buy now and also be charged a high price in the next period. It can also be shown under these conditions, and for the same reason, that the present value of profits is lower in this case than when the firm does not keep information about who are its previous customers. That is, in a monopoly setting, a firm knowing whom its previous customers are can have lower profits. As noted above, a variation of this possibility is when some consumers’ preferences change from this period to the next. In this case, the results above continue to go through, but we now get closer to the no information case. (p. 265) Another possibility is that there is entry of new consumers in the next period and the firm faces a market consisting of overlapping generations of consumers with each generation coming into the market in each period. In such a setting, a low price for new consumers satisfies all the existing low valuation consumers and may be followed in the next period by a high price for new consumers to take advantage of the new high valuation consumers coming into the market in the next period. Villas-Boas (2004) shows that this is indeed the case and that the equilibrium involves price cycles in both the prices to the new and previous customers. In periods when the firm charges low prices to new consumers, they know that the next period price is higher, and then they all buy in that period as long as their valuation is above the price charged. In periods when the firm charges high prices, consumers know that the next period prices to new consumers will be low, and therefore some consumers with valuation above the price charged decide not to buy. Another interesting issue to consider is what happens when consumers live as long as firms. Hart and Tirole (1988) consider this case when consumers can be of only two types. They find that if the discount factor δ 〉 1/2 there is no price discrimination when the horizon tends to infinity, with the firm charging always the low price. The intuition is that if the high valuation consumer ever reveals his valuation he will get zero surplus forever, and if there were an equilibrium where the high valuation consumer would buy at a price above the lowest price, he would gain from deviating and being seen as a low valuation consumer forever. Schmidt (1993) considers a variation of this setting where the seller is the one with private information (on its costs), with a discrete number of consumer types (possibly more than two types), and focusing on the Markov-perfect equilibria. The results there can also be presented in terms of private information by the consumers. In that case, the result is as in Hart and Tirole—there is Page 8 of 15 Price Discrimination in the Digital Economy no price discrimination as the horizon tends to infinity. The intuition is that if consumers are given a chance to build a reputation that they have a low valuation they will do so (see also Fudenberg and Levine 1989). Another interesting variation of randomly changing preferences is considered in Kennan (2001). There, a positive serial correlation leads to stochastic price cycles as purchases reveal a high consumer valuation, which is followed by a high price, while no purchases reveal low consumer valuation, which is followed by low prices. Another possibility discussed in the previous sections is the case where a firm when serving the customers that choose to purchase also learns their valuation. In that case consumers are also strategic in the current period, and the firm is hurt by the reduced number of consumers that decide to buy in the current period. However, because now the firm is able to extract in the next period all the valuation from the consumers that purchase in the current period, the present value of profits can actually be higher with this ability to learn the valuation of its previous customers. (p. 266) Competition We discuss now the effect of competition within the competition framework described in Section 2. When the firms are not able to recognize their previous consumers we are back to the no information case, and the market equilibrium is as characterized in Section 2 for that case. If firms are able to recognize their previous customers while noting that consumers that choose their product have a valuation above a certain threshold, we can use the analysis in the two previous sections to set up the problem of each firm. In order to present sharper results let us focus on the case of the uniform distribution of consumer valuations. The problem for Firm A can then be set as where v*(pfA, pfB) is the valuation for product A of the marginal consumer choosing product A in the current period, as computed in Section 3, and represents the next period profit for Firm A given the current period prices. For the uniform distribution, as we obtained in Section 3, we have Given this we can then write given v*. Note that . given the analysis of Section 2 of the equilibrium in the next period is composed of both the profits from Firm A‘s previous customers and the profits from Firm B's previous customers. In such a setting one can then show that the first period prices are higher than in the no information case. This is because the current period demand is less elastic than when there is no behavior-based price discrimination, as discussed in Section 3, because consumers know that if they do not switch the firm will offer them in the next period a lower price (a price to the new consumers in the next period). Note also that in this setting, and in relation to the symmetric outcome, if firms had more or fewer customers they would be better off in the next period. For the uniform distribution these two effects cancel out in the two-period model, and the first-period prices end up being independent of the firms’ discount factor (Fudenberg and Tirole 2000). If some of the consumer preferences can change from this period to the next, the results presented here would go through, but now the equilibrium prices would be lower and closer to the case where there cannot be behaviorbased price discrimination. Another interesting possibility is when there is entry of new consumers in each period in a set-up with overlapping generations of consumers, such that in each period new consumers for a firm could be either previous customers from the competitor or new consumers to the market. In this situation one has to fully consider the dynamic effects over several periods of the pricing decisions in the current period (Villas-Boas 1999). One result is that the prices to (p. 267) the new consumers end up being below the ones in the case with no behavior-based price discrimination if the consumers care enough about the future. This is because the demand from the new Page 9 of 15 Price Discrimination in the Digital Economy consumers is more elastic than in the case when no information is collected because the marginal consumers are just deciding the order in which they buy the different products, given that the marginal consumers switch products in equilibrium. Another possibility, along the lines discussed in the earlier sections, is when firms are able to learn the valuation of their previous consumers during the purchase process. In this case competing firms can potentially gain more from learning the valuation of consumers than not learning, which could be a force for greater competition for consumers. 5. Discussion and Related Topics The results above concentrate on the case where the firms can only offer short term contracts. In some cases firms may be able to attract consumers with the promise of guaranteed future prices. Hart and Tirole (1988) consider this situation in a monopoly setting, and show that with such contracts the seller is able to sell to the high valuation consumers at a price above the valuation of low valuation consumers, because it is able to commit to an average price in future periods that does not extract all the surplus from the high valuation consumers. Hart and Tirole also show that the outcome obtained is the same as if the seller were selling a durable good.3 This also shows that behavior-based price discrimination in a monopoly setting leads to a lower present value of profits than in the case of a monopolist selling a durable good. Battaglini (2005) considers the case of long-term contracts with two types of infinitely lived consumers with preferences changing over time (as in Kennan 2001) and continuous consumption. He shows that in the optimal contract the efficient quantity is supplied in finite time (after the type of the consumer is the high valuation type). Fudenberg and Tirole (2000) consider the effect of long-term contracts in competition in a two-period model. They show that long-term and short term contracts co-exist in equilibrium, with some consumers taking short-term contracts, and that there is less switching in equilibrium when long-term contracts are possible. The intuition for this last result is that the existence of long-term contracts (purchased by consumers with more extreme preferences) leads firms to be more aggressive (lower prices) in their short-term contracts, which yields less switching. Under competition and short-term contracts, Esteves (2010) and Chen and Zhang (2009) consider the case when the distribution of valuations is in discrete types. In this case, the equilibrium involves mixed strategies. Esteves considers the case with two consumer types, each having a preference for one of the competing firms, and myopic consumers, and finds that first period prices tend to fall when (p. 268) the discount factor increases. Chen and Zhang consider the case with three consumer types, with one type that has the same valuation for both firms, and two other types that have extreme preferences for each firm. In this set-up expected prices in the first period are higher even when consumers are myopic, as a firm that sells more in the first period is not able to discriminate in the next period between the consumer types that can consider buying from that firm. Another interesting issue to consider under competition is that in some markets where firms can practice behaviorbased price discrimination, consumers have also switching costs of changing supplier. Note that switching costs alone can lead to dynamic effects in the market as studied in the literature.4 The effects of the interaction of switching costs with behavior-based price discrimination are studied in Chen (1997) and Taylor (2003). Chen considers a two-period model and shows that with switching costs, behavior-based pricing leads to rising prices over time. Taylor considers the case of multiple periods and shows that prices are constant over time until the last period, and shows that moving from two firms to three firms, there is aggressive competition for the switcher consumers. In addition, Taylor considers the case when the firms may have some prior information about the switching costs of different consumers.5 On a related topic, firms could also offer different products depending on the purchase history of consumers. Some results on these effects are presented in Fudenberg and Tirole (1998), Ellison and Fudenberg (2000), and Zhang (2011). When studying the effects of price discrimination based on purchase history, one can also think of the implications for privacy concerns. For studies on the effects on privacy along these and related dimensions see, for example, Taylor (2004a, 2004b), Calzolari and Pavan (2006), and Ben-Shoham (2005).6 Page 10 of 15 Price Discrimination in the Digital Economy In some markets, firms also learn consumer characteristics that directly affect cost of servicing them. This is for example the case in credit markets, insurance markets, and labor markets. It is also the case when the costs of servicing customers are heterogeneous across consumers, and are learned through interaction with them. In the digital economy these aspects can become important with after-purchase service and the possibility of returns. The effects of these issues in credit markets are considered, for example, in Pagano and Jappelli (1993), Padilla and Pagano (1997, 2000), Dell’Ariccia et al. (1999), and Dell’Ariccia and Marquez (2004).7 Another important recent practice of firms with the digital economy is to make offers to consumers based on their search behavior. Armstrong and Zhou (2010) look at this case under competition and find that firms may have an incentive to offer a lower price on a first visit by a consumer than a return visit. This possibility can then lead to higher equilibrium prices in the market. For a recent survey of related effects of product and price comparison sites see Moraga and Wildenbeest's (2011) chapter 9 of this Handbook. Finally, another interesting issue to study would be the effect of purchase history on the advertising messages that consumers may receive.8 (p. 269) 6. Conclusion This paper presents a summary of the effects of price discrimination based on purchase history. With the digital economy this type of information became more available for firms, and consequently they are engaging more frequently in this type of price discrimination in markets such as telecommunications, magazine or newspaper subscriptions, banking services credit cards.9 For situations where firms have substantial market power (monopoly) we found that firms benefit after the information is gained, but that this may lead consumers to be more careful when revealing information, which could potentially hurt the firm's profits. For situations with competition, if the competitors are aware that firms have purchase history information, more information may actually lead to more intense competition after the information is gained. However, before information is gained, consumers may become less price sensitive as the marginal consumers may get a better price offer in the next period if they do not switch brands in response to a price cut. This may then lead to softer competition prior to firms gaining information. With new consumers coming into the market this effect is attenuated, as the prices for the new customers are now less affected by the fact that these new customers have a lower valuation for the firm. References Acquisti, A., Varian, H.R., 2005. Conditioning Prices on Purchase History. Marketing Science 24, 367–381. Armstrong, M., Zhou, J., 2010. Exploding Offers and Buy-Now Discounts. Working paper, University College London. Battaglini, M., 2005. Long-Term Contracting with Markovian Consumers. American Economic Review 95, 637–658. Belleflamme, P., Peitz, M., 2010. Industrial Organization: Markets and Strategies, Cambridge University Press: Cambridge, U.K. Ben-Shoham, A., 2005. Information and Order in Sequential Trade. Working paper, Harvard University. Borenstein, S., 1985. Price Discrimination in Free-Entry Markets. RAND Journal of Economics 16, 380–397. Bouckaert, J., Degryse, H., 2004. Softening Competition by Inducing Switching in Credit Markets. Journal of Industrial Economics 52, 27–52. Brandimarte, L., Acquisti, A., 2012. The Economics of Privacy. In M. Peitz, J. Waldfogel (eds.), The Oxford Handbook of the Digital Economy. Oxford/New York: Oxford University Press. Calzolari, G., Pavan, A., 2006. On the Optimality of Privacy in Sequential Contracting. Journal of Economic Theory 130, 168–204. Chen, Yongmin, 1997. Paying Customers to Switch. Journal of Economics and Management Strategy 6, 877–897. Page 11 of 15 Price Discrimination in the Digital Economy Chen, Yuxin, Iyer, G., 2002. Consumer Addressability and Customized Pricing. Marketing Science 21, 197–208. Chen, Yongmin, Pearcy, J.A., 2010. Dynamic Pricing: When to Entice Brand Switching and When to Reward Consumer Loyalty. RAND Journal of Economics 41, 674–685. Chen, Yuxin, Narasimhan, C., Zhang, Z.J., 2001. Individual Marketing and Imperfect Targetability. Marketing Science 20, 23–41. Chen, Yuxin, Zhang, Z.J., 2009. Dynamic Targeted Pricing with Strategic Consumers. International Journal of Industrial Organization 27, 43–50. Choi, J.P., 2012. Bundling Information Goods. In M. Peitz, J. Waldfogel (eds.), The Oxford Handbook of the Digital Economy. Oxford/New York: Oxford University Press. Corts, K.S., 1998. Third-Degree Price Discrimination in Oligopoly: All-Out Competition and Strategic Commitment. RAND Journal of Economics 29, 306–323. Dell’Ariccia, Friedman, E., Marquez, R., 1999. Adverse Selection as a Barrier to Entry in the Banking Industry. RAND Journal of Economics 30, pp. 515–534. (p. 271) Dell’Ariccia, Marquez, R., 2004. Information and Bank Credit Allocation. Journal of Financial Economics 72, 185– 214. Dobos, G., 2004. Poaching in the Presence of Switching Costs and Network Externalities. Working paper, University of Toulouse. Ellison, G., Fudenberg, D., 2000. The Neo-Luddite's Lament: Excessive Upgrades in the Software Industry. RAND Journal of Economics 31, 253–272. Engelbrecht-Wiggans, R., Milgrom, P., Weber, R., 1983. Competitive Bidding and Proprietary Information. Journal of Mathematical Economics 11, 161–169. Esteves, R.B., 2010. Pricing under Customer Recognition. International Journal of Industrial Organization 28, 669– 681. Farrell, J., Klemperer, P., 2007. Co-ordination and Lock-In: Competition with Switching Costs and Network Effects. In: M. Armstrong, R. Porter (Eds.), Handbook of Industrial Organization 3, Elsevier, pp. 1967–2072. Freixas, X., Guesnerie, R., Tirole, J., 1985. Planning under Incomplete Information and the Ratchet Effect. Review of Economic Studies 52, 173–191. Fudenberg, D., Levine, D., 1989. Reputation and Equilibrium Selection in Games with a Patient Player. Econometrica 57, pp. 759–778. Fudenberg, D., Tirole, J., 1998. Upgrades, Trade-ins, and Buybacks. RAND Journal of Economics 29, pp. 238–258. Fudenberg, D., Tirole, J., 2000. Customer Poaching and Brand Switching. RAND Journal of Economics 31, pp. 634– 657. Fudenberg, D., Villas-Boas, J.M., 2006. Behavior-Based Price Discrimination and Customer Recognition. In: T. Hendershott (Ed.) Handbooks in Information Systems: Economics and Information Systems, Chap.7, Elsevier, Amsterdam, The Netherlands. Hart, O.D. Tirole, J., 1988, Contract renegotiation and coasian dynamics, Review of Economic Studies 55, 509–540. Hermalin, B.E., Katz, M.L., 2006. Privacy, Property Rights & Efficiency: The Economics of Privacy as Secrecy. Quantitative Marketing and Economics 4, 209–239. Hirshleifer, J., 1971. The Private and Social Value of Information and the Reward to Inventive Activity. American Economic Review 61, 561–574. Page 12 of 15 Price Discrimination in the Digital Economy Holmes, T.J., 1989. The Effects of Third-Degree Price Discrimination in Oligopoly. American Economic Review 79, 244–250. Hui, K.-L., Png, I.P.L., 2006. The Economics of Privacy. In: T. Hendershott (Ed.) Handbooks in Information Systems: Economics and Information Systems, Elsevier, Amsterdam, The Netherlands. Iyer, G., Soberman, D., Villas-Boas, J.M., 2005. The Targeting of Advertising. Marketing Science 24, 461–476. Kennan, J., 2001. Repeated Bargaining with Persistent Private Information. Review of Economic Studies 68, 719–755. Moraga, J.-L., Wildenbeest, M., 2011. Price Dispersion and Price Search Engines. In M. Peitz, J. Waldfogel (eds.), The Oxford Handbook of the Digital Economy. Oxford/New York: Oxford University Press. Morgan, J., Stocken, P., 1998. The Effects of Business Risk on Audit Pricing. Review of Accounting Studies 3, pp. 365–385. Padilla, J., Pagano, M., 1997. Endogenous Communications among Lenders and Entrepreneurial Incentives. Review of Financial Studies 10, pp. 205–236. (p. 272) Padilla, J., Pagano, M., 2000. Sharing Default Information as a Borrower Incentive Device. European Economic Review 44, pp. 1951–1980. Pagano, M., Jappelli, T., 1993. Information Sharing in Credit Markets. Journal of Finance 48, pp. 1693–1718. Pazgal, A., Soberman, D., 2008. Behavior-Based Discrimination: Is It a Winning Play, and If So, When? Marketing Science 27, pp. 977–994. Roy, S., 2000. Strategic Segmentation of a Market. International Journal of Industrial Organization 18, pp. 1279– 1290. Schmidt, K., 1993. Commitment through Incomplete Information in a Simple Repeated Bargaining Game. Journal of Economic Theory 60, pp. 114–139. Shaffer, G., Zhang, Z.J., 2000. Pay to Switch or Pay to Stay: Preference-Based Price Discrimination in Markets with Switching Costs. Journal of Economics and Management Strategy 9, pp. 397–424. Sharpe, S., 1990. Asymmetric Information, Bank Lending and Implicit Contracts: A Stylized Model of Customer Relationships. Journal of Finance 45, 1069–1087. Shin, J., Sudhir, K., 2010. A Customer Management Dilemma: When Is It Profitable to Reward Existing Customers? Marketing Science 29, pp. 671–689. Stegeman, M., 1991. Advertising in Competitive Markets. American Economic Review 81, pp.210–223. Taylor, C.R., 2003. Supplier Surfing: Price-Discrimination in Markets with Repeat Purchases. RAND Journal of Economics 34, pp. 223–246. Taylor, C.R., 2004a. Consumer Privacy and the Market for Customer Information. RAND Journal of Economics 35, pp. 631–650. Taylor, C.R., 2004b. Privacy and Information Acquisition in Competitive Markets. Working paper, Duke University. Thisse, J.-F., Vives, X., 1988. On the Strategic Choice of Spatial Price Policy. American Economic Review 78, pp. 122–137. Ulph, D., Vulkan, N., 2007. Electronic Commerce, Price Discrimination and Mass Customization. Working paper, University of Oxford. Villanueva, J., Bhardwaj, P., Balasubramanian, S., Chen, Y., 2007. Customer Relationship Management in Competitive Environments: The Positive Implications of a Short-Term Focus. Quantitative Marketing and Economics 5, pp. 99–129. Page 13 of 15 Price Discrimination in the Digital Economy Villas-Boas, J.M., 1999. Dynamic Competition with Customer Recognition. RAND Journal of Economics 30, pp. 604– 631. Villas-Boas, J.M., 2004. Price Cycles in Markets with Customer Recognition. RAND Journal of Economics 35, pp. 486– 501. Wathieu, L., 2007. Marketing and the Privacy Concern. working paper, Harvard University. Zhang, J., 2011. The Perils of Behavior-Based Personalization. Marketing Science 30, pp. 170–186. Notes: (1.) Fudenberg and Villas-Boas (2006) discusses in further detail the market forces described here, and also discusses the effects of long-term contracts (including the relationship to durable goods and bargaining), multiple products and product design, switching costs, privacy concerns and credit markets. One important aspect of price discrimination in the digital economy that is not discussed here is that of bundling of information goods; see chapter 11 by Choi (2012) in this Handbook for a survey of this literature. For a textbook treatment of behaviorbased price discrimination see Belleflamme and Peitz (2010). (2.) The previous period consumer decisions are considered in the next section. (3.) Hart and Tirole also discuss what happens under commitment. See also Acquisti and Varian (2005) on the comparison of commitment with noncommitment. Acquisti and Varian consider also the effect of the seller offering enhanced services. (4.) See, for example, Farrell and Klemperer (2007) for a survey of the switching costs literature. (5.) For the analysis of the second period of a model with switching costs see Shaffer and Zhang (2000). Dobos (2004) considers the case of horizontal differentiation, switching costs, and network externalities. See also Villanueva et al. (2007) on the effect of customer loyalty, and Pazgal and Soberman (2008) on the incentives for firms in investing on technologies to track purchase histories. (6.) For related studies on privacy matters see also Hirshleifer (1971), Hermalin and Katz (2006), and Wathieu (2007). See also chapter 20 by Brandimarte and Acquisti (2012) in this Handbook for a survey on the economics of privacy. For a recent survey on privacy issues related to information technology see Hui and Png (2006). (7.) See also Engelbrecht-Wiggans et al. (1983), Sharpe (1990), Morgan and Stocken (1998), and Bouckaert and Degryse (2004). (8.) For studies of targeted advertising see, for example, Stegeman (1991), Roy (2000), and Iyer et al. (2005). Also related is the literature on competition with price discrimination such as Thisse and Vives (1988), Borenstein (1985), Holmes (1989), Corts (1998), Chen et al. (2001), Chen and Iyer (2002), and Ulph and Vulkan (2007). (9.) See for example, “Publications are Trying New Techniques to Win over Loyal Readers.” The New York Times, January 4, 1999, p. C20. Drew Fudenberg Drew Fudenberg is the Frederick E. Abbe Professor of Economics at Harvard University. J. Miguel Villas-Boas J. Miguel Villas-Boas is the J. Gary Shansby Professor of Marketing Strategy at the Haas School of Business, University of California, Berkeley. Page 14 of 15 Bundling Information Goods Oxford Handbooks Online Bundling Information Goods Jay Pil Choi The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0011 Abstract and Keywords This article, which reports the theory of product bundling, is particularly relevant in the context of the digital economy. It also addresses the price discrimination theory of bundling. Firm's bundling decision may alter the nature of competition and thus have strategic effects in its competition against its rivals. In the case of competitive bundling, the strategic incentives to bundle depend on the nature of available products and the prevailing market structures for the available products. The incentives to bundle depend on the influences of bundling on price competition. Consumers indirectly benefit from the number of adopters of the same hardware. Microsoft's bundling practices have found antitrust investigations on both sides of the Atlantic. Bundling is a profitable and credible strategy in that it increases the bundling firm's profit and may make the rival firms unable to sell their journals. The leverage theory of bundling has significant implications for antitrust analysis. Keywords: product bundling, digital economy, price discrimination, firms, competition, competitive bundling, Microsoft, leverage theory, antitrust analysis 1. Introduction Bundling is a marketing practice that sells two or more products or services as a package. For instance, the socalled “Triple Play” bundle in the telecommunication sector combines phone, Internet, and TV services, which is often offered at a discount from the sum of the prices if they were all bought separately. As more people adopt ereader devices such as Amazon Kindle and Barnes & Nobel Nook, publishers are increasingly offering bundles of digital titles to boost demand, and the bundling strategy is expected to grow in popularity as sales of electronic books surge.1 Sometimes, bundling takes the form of technical integration of previously separate products with the evolution and convergence of technologies. For instance, smartphones provide not only basic phone services, but also offer advanced computing abilities and Internet connectivity as well as other features such as digital cameras, media players, and GPS capabilities, to name a few, which were all offered previously by separate products . Computer operating systems exhibit a similar pattern over time. Modern operating systems such as Microsoft's Windows 7 and Apple's Mac OS X Snow Leopard include numerous features that were not part of earlier operating systems. Many of these features originally were offered only as separate software products. Economists have proposed many different rationales concerning why firms practice bundling arrangements as a marketing tool. This chapter reviews the literature on bundling and suggests topics for future research with special emphasis on information goods. In particular, it discusses the nature of information goods (p. 274) and how the special characteristics of information goods make the practice of bundling an attractive marketing strategy in the digital economy. There are two types of bundling. With pure bundling, a firm offers the goods for sale only as a bundle. With mixed bundling, a firm also offers at least some individual products separately in addition to the bundle, with a price for the bundle discounted from the sum of the individual good prices. In the monopoly setting without any strategic interaction, we can immediately rule out pure bundling as a uniquely optimal strategy Page 1 of 23 Bundling Information Goods because pure bundling can be considered as a special case of mixed bundling in which individual products are offered at the price of infinity. However, in the context of strategic interaction, pure bundling can be used as a commitment device and be a uniquely optimal strategy, as will be shown. There are many reasons that firms offer bundled products. One obvious reason is efficiency due to transaction costs. Technically speaking, any reasonably complex product can be considered as bundled products. For instance, a car can be considered as a bundle of engine, car body, tires, and so on. Obviously, it is much more efficient for the car manufacturers to assemble all the parts and sell the “bundled” product rather than for consumers to buy each part separately and assemble the car themselves. Other motives for bundling identified in the economics literature include reducing search and sorting costs (Kenney and Klein, 1983), cheating on a cartel price, evasion of price controls, protection of goodwill reputation, and so on.2 In this chapter, I focus on three most prominent theories of bundling, which are price discrimination, competitive bundling, and the leverage theory, with special emphasis on information goods. In particular, I discuss special features of information goods and their implications for the bundling strategy. First, information goods have very low marginal costs (close to zero) even though the costs of producing the first unit might be substantial. The low marginal costs of information goods make bundling a large number of goods an attractive marketing strategy even if consumers do not use many of them. For instance, major computer operating systems today include a host of functionalities and features most consumers are not even aware of, let alone use. As I discuss later, this property has important implications for the price discrimination theory of bundling. Second, information goods are often characterized by network effects, meaning that the benefits a user derives from a particular product increases with the number of other consumers using the same product. For example, the utility of adopting particular software increases as more consumers adopt the same software because there are more people who can share files. A larger user base also induces better services and more development of complementary products that can be used together. With network effects, bundling can be a very effective leverage mechanism through which a dominant firm in one market can extend its monopoly power to adjacent markets that otherwise would be competitive. Thus, network effects can be an important factor in the bundling of information goods with strategic reasons. Finally, it is worthwhile mentioning that the convergence in digital technologies also facilitates bundling of diverse products which were previously considered (p. 275) as separate. Consider, for instance, the convergence of broadcasting and telephone industries. Traditionally, they represented very different forms of communications in many dimensions including the mode of transmission and the nature of communication. However, the convergence of digital media and the emergence of new technologies such as the Internet have blurred the boundaries of the previously separate information and communication industries because all content including voice, video, and data can now be processed, stored and transmitted digitally (Yoo, 2009). As a result, both telephone companies and cable operators can offer bundled services of phone, Internet, and TV and compete head-to-head. The rest of the chapter is organized as follows. In Section 2, I start with the price discrimination theory of bundling with two goods (Stigler, 1963) and extend the model to study the strategy of bundling a large number of information goods that is based on the law of large numbers (Bakos and Brynjolfsson, 1999). In Section 3, I review strategic theory of bundling and discuss the relevance of these theories for information goods and how they can be applied to specific cases. I first discuss competitive bundling in which entry and exit is not an issue in the bundling decision. Then, I move on to the leverage theory of bundling in which the main motive is to deter entry or induce the exit of rival firms in the competitive segment of the market. In the review of the leverage theory of bundling, I also explore antitrust implications and discuss prominent cases, as well as investigating optimal bundling strategies and their welfare implications. Throughout the chapter, I aim to discuss how the predictions of the model match with empirical observations and evidence. 2. Bundling Information Goods for Price Discrimination 2.1. Price Discrimination Theory of Bundling The idea that bundling can be used as a price discrimination device was first proposed by Stigler (1963). The Page 2 of 23 Bundling Information Goods intuition is that bundling serves as a mechanism to homogenize consumer preferences over a bundle of goods under certain conditions, which facilitate extraction of consumer surplus. To understand the mechanism, consider the following example. Suppose that there are two products, Word Processor and Spreadsheet. For each product, consumers either buy one unit or none. Assume that there is no production cost. There are two types of consumers, English Major and Finance Major. The total number of consumers is normalized to 1, with the mass of each type being 1/2. The reservation values for each product by different types of consumers are given by the Table 11.1. (p. 276) Table 11.1 Word Processor Spreadsheet English Major $60 $40 Finance Major $40 $60 It can be easily verified that if Word Processor and Spreadsheet are sold separately, the optimal price for each product is $40 and all consumers buy the product. Thus, the total profit for the firm would be $80 (=$40 + $40). Now let us assume that the firm sells the two products as a bundle.3 Then, both types of consumers are willing to pay up to $100 for the bundle. As a result, the firm can charge $100 and receives a profit of $100 from the bundle, which is more profitable than selling the two products separately. Notice that both types of consumers pay the same price for the bundle, but effectively they pay different prices for each product. To see this, we can do the thought experiment of how much each consumer is paying for each product. English Majors implicitly pay $60 for the Word Processor and $40 for the Spreadsheet program. In contrast, Finance Majors implicitly pay $40 for the Word Processor but $60 for the Spreadsheet. In this sense, bundling can be considered as a price discrimination device.4 As illustrated by the example, if consumers have negative correlation of reservation values across products, bundling will homogenize consumers’ total reservation value for the collection of the products. This was first pointed out by Stigler (1963) and the idea has been extended by Adams and Yellen (1976) who compare the relative merits of independent pricing, mixed bundling and pure bundling. From the perspective of an optimal pricing mechanism, it is clear that mixed bundling (weakly) dominates independent pricing and pure bundling because the latter two can be considered as special cases of mixed bundling.5 Their paper thus focuses on under what conditions mixed bundling can strictly dominate the other two. To understand those conditions, Adams and Yellen list three desiderata that are satisfied by perfect price discrimination and can be benchmarks against which the profitability of other pricing schemes can be measured: 1. Extraction: No individual realizes any consumer surplus on his purchases. 2. Exclusion: No individual consumes a good if the cost of that good exceeds his reservation price for it. 3. Inclusion: Any individual consumes a good if the cost of that good is less than his reservation price for it. The relative merits of each pricing scheme depend on how well these desiderata are satisfied. For instance, independent pricing always satisfies Exclusion because individual prices are never set below cost. However, the other two criteria (Extraction or Inclusion) are easily violated by independent pricing, but can be mitigated by pure bundling if consumers’ preferences are negatively correlated, and as a result bundling reduces consumer heterogeneity. One major drawback of (p. 277) pure bundling is in complying with Exclusion. This can be a serious problem if the cost of each good is high and there is a substantial chance that the reservation value of consumers can be less than the production cost. In such a case, mixed bundling can be a way to mitigate the problem. To see this, consider the following numerical example. There are four consumers whose reservation values for two goods A and B are given, respectively, by (10, 90), (40, 60), (60, 40), and (90, 10). Let the production cost of each good be 20. It can be easily verified that the optimal prices under an individual pricing scheme are given by pA* = pB*=60 with an overall profit of 160. In this example, consumers’ reservation values are perfectly negatively Page 3 of 23 Bundling Information Goods correlated with the identical reservation value of 100 for the bundle, which allows the Extraction and Inclusion requirements to be simultaneously satisfied by pure bundling. The optimal pure bundle price is P*=100 with a profit of 240, which is higher than the profit under an individual pricing scheme. However, pure bundling in this example entails inefficiency due to the violation of Exclusion in that the first consumer purchases good A although his reservation value for the good (10) is less than its production cost (20). Similarly, the fourth consumer inefficiently purchases good B under a pure bundling scheme. This problem can be avoided by the use of mixed bundling in which the bundle is priced at 100 as before, but individual components are priced at pA* = pB*=90 −εεε, which induces these two consumers to purchase only good B and A, respectively. The resulting profit from mixed bundling is 260. The additional profit of 20 compared to the pure bundling scheme comes from the avoidance of wasteful consumption with pure bundling. To foreshadow our discussion later, note that the major benefit of mixed bundling over pure bundling comes from the mitigation of the Exclusion problem associated with pure bundling, which arises when the production cost of each good is positive. For information goods, the marginal cost is almost negligible and Exclusion is not a serious issue. This implies that pure bundling is as good as mixed bundling with information goods, as elaborated below. The early analyses of bundling, exemplified by Stigler and Adams and Yellen, were based on a series of numerical examples and left an impression that negative correlation is required for the bundle to be profitable. Schmalensee (1984) extends the analysis to the case where consumers’ reservation values are drawn from bivariate normal distributions and show that pure bundling can be more profitable even when the correlation of consumers’ valuations is nonnegative. McAfee, McMillan, and Whinston (1989; hereafter MMW) further expand the analysis by providing the first systematic approach to find conditions under which (mixed) bundling can be profitable. In particular, they show that mixed bundling is profitable with independent reservation values. As a simple illustration of this, consider an example in which individuals have the following four valuations with equal probability—(40,40), (40,70), (70,40), (70,70)—and no costs. This example satisfies the independence assumption. It can be easily verified that selling individual products yields the profit of 320 while selling a bundle yields a profit of 330. To derive a sufficient condition, MMW conduct the following heuristics. Suppose that pA* and pB* denote the optimal prices for goods A and B, respectively, (p. 278) when they are sold independently. Consider an alternative strategy of mixed bundling in which the bundle is priced at pA* + pB*, while still offering individual products at the same price as before. Then it is clear that this alternative scheme has no effect on consumers’ purchase patterns and the firm's profits. Now suppose that one of the prices for the individual goods, say B, is increased by ε, that is, the two goods are also offered independently at pA* and pB* + ε. There are three first order effects of locally raising pB in the mixed bundling scheme. First, there is a direct price effect from the inframarginal consumers who buy only good B, which has a positive impact on the firm's profit. Second, there is reduction in demand for good B due to the loss of marginal consumers who buy only good B, which has a negative effect on the firm's profit. Third, sales of good for A increase due to consumers who switch from buying only good B to purchasing the bundle. This last effect is positive. The considerations of these three effects yield the sufficient condition identified in MMW. The condition is the first general result that allows for arbitrary joint distributions and constitutes a significant advance over the previous literature that has relied mostly on numerical examples and simulations. One corollary that comes out of MMW's approach is that mixed bundling dominates independent sales when consumers’ reservation values are independently distributed. To see this, consider the first two effects of a local change in the price of B. These two effects are exactly the same considerations the monopolist would consider in setting its price if it were selling good B to consumers whose valuations of good A is less than pA*. With the independence assumption, however, the consumers’ demand for good B is independent of their valuations for good A, which implies that the first two effects must cancel out and be zero at the optimal price pB* by definition. Thus, only the third effect remains and mixed bundling is profitable with independent distribution of reservation values across products. The sufficient condition identified in MMW suggests that mixed bundling is optimal in a wider range of cases than just the independent distribution case. However, the condition is somewhat difficult to interpret. More recently, Chen and Riordan (2010) use a new statistical concept called copulas to model stochastic dependence of consumers’ reservation values.6 The use of copulas yields new analytical results that are easier to interpret. They strengthen the result of MMW to show that a multiproduct firm achieves a higher profit from mixed bundling than separate selling if consumer values for the two products are negatively dependent, independent, or have limited 7 Page 4 of 23 Bundling Information Goods positive dependence.7 Next, I further discuss the relevance of their results in relation to information goods. 2.2. Price Discrimination Theory of Bundling Applied to Information Goods In this section, we apply the price discrimination theory of bundling to information goods. To this purpose, there are two important characteristics of information goods that are relevant and need to be taken into account. (p. 279) First, as pointed out by Arrow (1962), information has the property of a public good in that the consumption of information by one agent does not affect the availability of that good to other potential users; once the first unit is produced, marginal cost for additional units is essentially zero. Second, information goods can reside in cyberspace and need not be limited by physical limitations. As a result, a large number of information goods can be bundled without incurring any substantial costs. Note that, in contrast to information goods, with some conventional goods there is a non-trivial cost to consumers of including unwanted features, over and above any impacts on price. To appreciate the absence of physical limitation in the bundling decision, consider the example of Swiss army knives. They typically include various tools in addition to knife blades and are offered in many different configurations and prices. Adding additional tools not only raises the marginal cost of manufacturing, but it also makes the knife bulkier, which most users dislike (all else equal). As a result, manufacturers need to offer a wide variety of types of Swiss army knives that vary in both the number and mix of tools, thus catering to different consumer preferences. A knife with all possible tools would be both expensive to produce and too bulky for most consumers. In contrast, a bundle of information goods that includes far more products than any one consumer wants can appeal to a large number of potential buyers (who vary in which subset of the bundled products they will use, but can ignore the rest) and can be much less expensive to produce than a series of different versions, each with a limited set of products that appeals to a particular subset of consumers. As a result, we observe the providers of information goods engage in large scale bundling of information goods. For instance, cable television companies typically bundle hundreds of channels, and major publishers bundle a large number of their e-journals. The use of Internet as a distribution channel of digital information also reduces distribution costs and facilitates large scale bundling. The implications of these characteristics of information goods for bundling have been investigated by Bakos and Brynjolfsson (1999).8 More specifically, to reflect the typical cost and demand conditions for information goods, assume that the marginal cost for copies of all information goods is zero to the seller and each buyer has unit demands for each information good. In addition, assume that buyer valuations for each product are independent and identically distributed (i.i.d.) with a finite mean (μ) and variance (σ2 ).9 Under these assumptions, Bakos and Brynjolfsson find that selling a bundle of information goods can be vastly superior to selling the goods separately. In particular, they show that for the distributions of valuations underlying many common demand functions, bundling substantially reduces average deadweight loss and leads to higher profits for the seller. They derive an asymptotic result that as the number of products included in the bundle (denoted as N) increases, the deadweight loss per good and the consumers’ surplus per good for a bundle of N information goods converges to zero, and the seller's profit per good is maximized, approaching the level achieved under perfect price discrimination. (p. 280) The intuition for this result comes from the law of large numbers. As N increases, the intuition of Stigler operates with a vengeance. As the number of products included in the bundle increases without bound, the average valuation of each product in the bundle converges to the mean value μ. As a result, the seller can capture an increasing fraction of the total area under the demand curve, correspondingly reducing both the deadweight loss and consumers’ surplus relative to selling the goods separately. The idea can be graphically illustrated with the following example in Bakos and Brynjolfsson (1999).10 Assume that each consumer's valuations for individual goods are drawn from a uniform distribution on [0,1]. Let the number of consumers be normalized to 1. Then, the demand curve for each individual good is given by a linear inverse demand curve P=1−Q. As the bundle size increases, the per good inverse demand curve can be represented in Figure 11.1. As can be seen from Figure 11.1, as the bundle size increases, the law of large numbers assures that the distribution for the valuation of the per good included in the bundle coalesces around the mean value (μ) of the underlying distribution, yielding a per good demand curve that becomes more elastic around the mean and less Page 5 of 23 Bundling Information Goods elastic away from it. As a result, the seller can capture the most area under the demand curve, reducing both the deadweight loss and consumers’ surplus compared to the individual pricing scheme. Click to view larger Figure 11.1 The Per Good Inverse Demand Curve for Different Bundle Sizes. N=1, 2, and 20. Two remarks are in order concerning the profitability of bundling information goods. First, one implicit assumption in the discussion of large scale bundling of information goods is that there is no disutility associated with it from the perspective of consumers. Certainly, the cost is much smaller compared to physical goods as discussed earlier, but there still could be non-negligible costs for consumers to navigate in the set of offerings or to customize the bundled offer to their own preferences, especially if the bundling is on a very large scale and creates the problem of “information overload.”11 This type of diseconomies on the demand side may limit the size of bundles and firms may need to provide customized bundles tailored to consumers’ preferences, which is made possible with the availability of (p. 281) massive consumer data and sophisticated data mining technologies, as discussed further below. Second, with the development of the Internet technology and electronic commerce, firms can easily trace and store consumers’ records of previous purchases by electronic “fingerprints.” They can use this information to infer consumers’ preferences, which enable them to practice mass customization that produces goods tailored to each customer's needs without significantly sacrificing cost efficiency.12 This development has the potential to obviate the need to use bundling as a price discrimination device. The attractiveness of mass customization is especially attractive for physical goods with substantial marginal costs. However, firms also have ex post incentives to use the information gleaned from past sales record to engage in the so-called behavior-based price discrimination.13 In response to such incentives, consumers can engage in “strategic demand reduction” to misrepresent their preferences to elicit a lower price in the future. This strategic response by consumers can hurt the firms’ ex ante profits.14 Bundling avoids such pitfalls and can be considered as a commitment mechanism not to engage in personalized pricing in the future. In addition, for information goods there is less concern for the inefficiency of bundling that may arise by forcing consumers to purchase goods that they value below marginal cost. Thus, the importance of bundling as a pricing strategy is expected to continue for information goods even in the new environment. Chen and Riordan (2010) also are worthy to mention in relation to information goods. As discussed earlier, they provide analytical results that mixed bundling yields higher profits than separate selling if consumer values for the two products are negatively dependent, independent, or have limited positive dependence. The reason they need some restriction on the degree of positive dependence is that if consumers’ reservation values of the products are perfectly positively dependent, bundling yields the same profit as separate selling. The exact condition they derive for the profitability of mixed bundling shows that the restriction is relaxed as the marginal costs for the two products are lowered. This implies that mixed bundling is more likely to be profitable for information goods with zero marginal costs, with all other things being equal. In addition, Chen and Riordan (2010) extend their analysis of product bundling for 2 goods to a more general ngood case, and consider a more general pricing scheme than pure bundling but much simpler and practically easier to implement than mixed bundling, which they call step bundling (SB). We know that mixed bundling is (weakly) superior to pure bundling as a pricing scheme. However, as the number of goods included in the bundle increases, the complexity of the mixed bundling scheme exponentially worsens. If there are n products to be included in the bundle, there are (2n −1) possible combinations of products and corresponding prices. For instance, if there are 10 products to bundle, there are 1023 prices that need to be set for full implementation of a mixed bundling scheme. As a result, mixed bundling becomes quickly impractical with the increase of products in the bundle. This explains why Bakos and Brynjolfsson's analysis is focused on pure bundling when they consider bundling of a large number of information goods. Page 6 of 23 Bundling Information Goods (p. 282) Chen and Riordan propose a much simpler scheme. Let N be the set of products offered by a firm. In a kstep bundling scheme, a firm offers individual prices, pi, for each good i∈N, and bundle prices Pl for any bundles that include l goods from Nl, where Nl, ∈ N, l = 1, 2,…, k, and k ∈ [2, |N|]. SB may include part or all k-step bundling. As an example, consider a cable company that offers phone, Internet, and TV services. Each service can be offered with individual prices that can differ across services. In addition, it also offers a 2-good bundle that include Internet and TV services, and the Triple Play bundle that includes all three services. This would be an example of 3-step bundling. Chen and Riordan extend their analysis of product bundling for 2 goods to a more general n-good case and show that their main results on the profitability of mixed bundling naturally extends to kstep bundling. Chu, Leslie, and Sorensen (2011) propose a further simplified form of mixed bundling, called bundle-size pricing (BSP). In this pricing scheme, a firm sets different prices for different sized bundles. There is one price for any single good. There is another price for any combinations of two goods, and so on. With such a pricing mechanism, there are only N prices if there are N products that constitute the “grand” bundle. So if there are 10 products to bundle, this requires only 10 prices, compared to 1023 prices with the mixed bundling scheme. BSP differs from SB in that there is only one price for all individual goods there is no restriction on what goods are allowed to be included in a bundle. More precisely, SB is identical to BSP if pi = p for all i∈N, and Nl = N, for all l = 2,…, |N|. Thus, BSP can be considered as a special form SB. Chu et al. show that BSP attains almost the same profit as mixed bundling under most circumstances. The main intuition for their result, once again, comes from the heterogeneityreduction effect of bundling. If bundles are large, different bundles of the same size do not need to be priced very differently as the heterogeneity of consumers is reduced with bundling. As a result, prices for large-sized bundles under bundle-size pricing tend to be very close to those under mixed bundling. The price pattern under BSP can be very different for individual goods, but it turns out that bundles are much more important to profits and the discrepancies in individual good prices play a negligible role. This implies that BSP approximates mixed bundling in terms of profits. By extension, this also implies that SB attains almost the same profit as mixed bundling under most circumstances. Many of the prices under mixed bundling thus are redundant. In contrast to Bakos and Brynyolfsson who derive asymptotic results as N→∞, the analyses of Chen and Riordan (2010) and Chu et al. (2011) focus on a finite N. With numerical analysis, Chu et al. show that in most cases BSP attain nearly the same level of profits attainable under mixed bundling. To illustrate the empirical relevance of their findings, they also estimate the demand facing a theater company that produces a season of 8 plays and compute the profitability of various pricing schemes. They show that bundle-size pricing is 0.9 percent more profitable than individual pricing and attains 98.5 percent of the mixed bundling profits. In this particular case, we can argue that bundling does not confer significant benefits over component pricing.15 They attribute the small effect of BSP to a very (p. 283) high degree of positive correlation in the consumers’ preferences and the empirical results may understate the gains from BSP in other settings. Their study may have important practical implications for a small to moderate number of products. If the number of products becomes large, Bakos and Brynyolfsson show that pure bundling approximates perfect price discrimination and thus BSP does not confer any significant advantage over pure bundling. Shiller and Waldfogel (2011) is a rare empirical study that explores that profit and welfare implications of various pricing scheme including bundling for one type of information goods. Based on survey data on about 1000 students’ valuations of 100 popular songs in early 2008 and 2009, they estimate how much additional profit or consumer surplus can be achieved in alternative pricing schemes vis-à-vis uniform pricing in the context of digital music. Their study was motivated by the Apple iTune Music Store's puzzling practice of charging a uniform price of $0.99 for all songs, regardless of their popularity. Even though their samples of songs and survey respondents are not representative, their empirical result illustrates the potential profitability of bundling strategies for information goods. In particular, their estimation suggests that the optimal pure bundling prices for 50 songs raise revenues by 17 percent and 29 percent, respectively, relative to uniform pricing with the 2008 and the 2009 samples. These gains come partly at the expense of consumers. Consumer surpluses under pure bundling are reduced by 15 percent and 5 percent, respectively, in years 2008 and 2009, compared to those under uniform pricing. They also show that the benefit of bundling increases with the bundle size, but at a decreasing rate; more than half of the benefit from the 50-song bundle is achieved with bundles of just 5. Finally, Bakos and Brynjolfsson (2000) extend their monopolistic analysis to different types of competition, including both upstream and downstream, as well as competition between a bundler and a single good producer and Page 7 of 23 Bundling Information Goods competition between two bundlers of different sizes. They show that bundling confers significant competitive advantages to bundling firms in their competition against non-bundlers due to their ability to capture a larger share of profits compared to smaller rivals.16 In the next section, we delve into the analysis of bundling as a strategic weapon against rival firms in more detail. 3. Bundling Information Goods for Strategic Reasons In section II, we considered price discrimination motives of bundling in the monopoly context. In oligopolistic markets, firms may have incentives to bundle their products for strategic reasons. More specifically, a firm's bundling decision may change (p. 284) the nature of competition and thus have strategic effects in its competition against its rivals. In the presence of (potential) competitors, the strategic effects of bundling thus should be considered in conjunction with any price discrimination motives. I first provide an overview of strategic theory of bundling and discuss the relevance of these theories for information goods and how they can be applied to specific cases. In the discussion of the theory, I distinguish two cases depending on whether bundling is used as an exclusionary strategy: competitive bundling and the leverage theory. The case of competitive bundling considers a situation in which the rivals’ entry or exit decisions are not affected by bundling and thus exclusion does not take place. As shown below, in the case of competitive bundling, the strategic incentives to bundle depend crucially on the nature of available products and the prevailing market structures for the available products. Bundling arrangements can also be used as an exclusionary device and have potential to influence the market structure by deterring entry or inducing exit of the existing firms. This issue is addressed in the discussion of the leverage theory of bundling. 3.1. Competitive Bundling The models of competitive bundling consider a situation in which firms compete, but do not intend to drive rival firms out of the market because the strategy of market foreclosure is too costly or simply not possible. In such a scenario, it is every firm's interest to soften price competition in the market. The incentives to bundle thus depend on the effects of bundling on price competition. 3.1.1. Bundling as a Product Differentiation Strategy When firms produce a homogenous product and compete in prices as in the standard Bertrand model, bundling can be used as a product differentiation strategy and can relax price competition. Chen (1997), for instance, considers a duopoly competing in the primary market and prefect competition prevailing in the production of other goods with which the primary good can be bundled. He considers a two-stage game in which the two firms in the primary market decide whether to bundle or not, and then they compete in prices given their bundling strategies chosen in the previous stage. He shows that at least one firm in the duopoly market chooses the bundling strategy in equilibrium, and both firms earn positive profits even though they produce a homogenous product and compete in prices.17 The intuition behind this result is that the bundled product is differentiated from the stand-alone products and allows above-cost pricing with market segmentation.18 Carbajo, De Meza, and Seidman (1990) provide a related model in which bundling may be profitable because it relaxes price competition. More specifically, they consider a situation in which one firm has a monopoly over one good but competes with another firm in the other market. In addition, they assume a prefect positive (p. 285) correlation between the consumers’ reservations prices of the two goods to eliminate any bundling motives that come from price discrimination. In their model, bundling provides a partitioning mechanism to sort consumers into groups with different reservation price characteristics. The firm with bundled products sells to high reservation value consumers while the competitor sells to low reservation value consumers. In contrast to Chen (1997) where bundling relaxes price competition in the primary (tying good) market, bundling in Carbajo et al. relaxes price competition in the secondary (tied good) market. In particular, if Bertrand competition prevails in the duopoly market, bundling is always profitable in their model.19 Bundling, once again, provides a way to differentiate their product offerings and soften price competition. 3.1.2. Bundling in Complementary Markets Page 8 of 23 Bundling Information Goods Consider a situation in which individual products, A and B, which can be included in the bundle, are perfect complements that need to be used on a one-to-one basis. For instance, they can be considered as component products that form a system. There are two firms, 1 and 2, in the market, whose unit production costs for components A and B are denoted by ai and bi, respectively, where i = 1, 2. For simplicity, assume that the two firms produce homogenous products and consumers have an identical reservation value for the system good. In this setup, it is easy to demonstrate that unbundling always (weakly) dominates a bundling strategy. If the two products are not bundled, the consumers can mix and match and will buy each component from the vendor who charged the lowest price so long as the sum of the two prices does not exceed the reservation value, which is assumed to be sufficiently large. Without bundling, competition will be at the component level and in each component market the firm with the lower production cost will win the whole market by selling at the production cost of the rival firm (minus ε). The prices in component markets A and B will be realized at max (a1, a2) and the max (b1, b2), respectively. If two products are bundled, the two products must be purchased from the same firm. In other words, bundling changes the nature of competition from localized competition to all-out competition. As a result, the firm with the overall advantage in production costs will make all sales at the system cost of the rival firm. The total price consumers pay for the bundled system is max (a1+ b1, a2 + b2), which is equal to or less than [max (a1, a2) + max (b1, b2)], the total price to be paid if the two components are sold separately. In this simple setup, there is no difference between bundling and no bundling if one firm is more efficient in the production of both components. However, if one firm is more efficient in one component and the other firm is in the other, unbundling is more profitable for the firms. To see this, without loss of generality, assume that a1〈 a2, b1 〉b2, and a1+ b1 〈 a2 + b2, that is, firm 1 is more efficient in component A and in the overall components while firm 2 is more efficient in component B. Then, each firm's profit under unbundling is given by (p. 286) (a2 −a1) 〉 0 and (b1 − b2 ) 〉0 for firm 1 and firm 2, respectively. In contrast, the corresponding profits under bundling is given by (a2 + b2 ) – (a1+ b1) and zero, respectively for firm 1 and firm 2. Since (a2 + b2) − (a1 + b1) 〈 (a2 − a1) under our assumption, unbundling is more profitable for both firms. Matutes and Regibeau (1988) develop this idea further in a model of complementary products with product differentiation and heterogeneous consumers. Consumers’ preferences are represented by a point in a twodimensional Hotelling model. They show that the pure bundling strategy is dominated by the individual marketing strategy. In their model with product differentiation, there can be an added advantage associated with unbundling. With bundling, there are only two systems available. With unbundling, consumers can mix and match and form four different systems, which enables consumers to choose systems that are closer to their ideal specifications. This implies that in system markets, firms would be reluctant to engage in competitive bundling unless there are other countervailing incentives to bundle. 3.1.3. Bundling in Substitute Markets and Inter-Firm Joint Purchase Discounts Armstrong (2011) extends the bundling literature in two respects by allowing products to be substitutes and supplied by separate sellers.20 An example of bundling in substitute markets is a “city pass” that allows tourists to visit all attractions in a city. However, tourists are typically time-constrained and can visit only a few during their visits. He shows that firms have incentives to engage in bundling or joint purchase discounts when demand for the bundle is elastic relative to demand for stand-alone products. This result holds regardless of whether component products are produced by one integrated firm or separate firms. When products are supplied by separate firms, Armstrong (2011) finds that they can use joint purchase discounts as a mechanism to relax price competition. The reason is that when firms offer a joint purchase discount, the innate substitutability of their products is reduced, which enables them to set higher prices. This implies that the seemingly consumer-friendly joint purchase discount scheme can be used as a collusive mechanism, and thus should be viewed with a skepticism by antitrust authorities. He also shows that a firm often has a unilateral incentive to offer a joint-purchase discount when their customers buy rival products. In such a case, joint purchase discounts can be implemented without any need for coordination between firms. In practice, the implementation of joint purchase discounts by separate firms can be difficult due to Page 9 of 23 Bundling Information Goods the need to make sequential purchases and produce a verification of purchase from the rival firm. These problems can now be mitigated by online platforms or product aggregators that allow simultaneous purchases. 3.2. The Leverage Theory of Bundling According to the leverage theory of bundling, a multiproduct firm with monopoly power in one market can monopolize a second market using the leverage provided by its monopoly power in the first market. This theory, however, has been controversial. I briefly describe the evolution of this theory. (p. 287) 3.2.1. One Monopoly Profit and the Chicago School Critique At first, it may seem that anticompetitive bundling will be profitable for a monopolist in many circumstances because monopolists can “force” consumers to take what they do not want in exchange for being allowed to purchase the monopolized product. Under that view, by bundling the monopolized product with otherwise competitive products, the monopolist can drive others from the adjacent markets, and thus achieve power in those markets. However, the logic of the theory has been criticized and subsequently dismissed by a number of authors from the University of Chicago school such as Bowman (1957), Posner (1976), and Bork (1978) who have argued that the use of leverage to affect the market structure of the tied good (second) market is impossible. In particular, the Chicago school critique points out that that monopolists already have economic profits in the market they monopolize, and there are costs to anticompetitive bundling. A monopolist does not have the freedom to raise prices or impose burdensome conditions on consumers without suffering reduced sales. In reality, the monopolist's pricing decision is constrained by market demand, that is, consumers’ willingness to pay for the product, which depends in part on the availability of substitutes. As a result, an enterprise, even a monopolist, foregoes profits in its original market when it engages in bundling that does not generate efficiencies and reduces consumer value. Such anticompetitive bundling leads to lower demand for the monopolist's product. To quote Posner (1976), “let a purchaser of data processing be willing to pay up to $1 per unit of computation, requiring the use of 1 second of machine time and 10 punch cards, each of which costs 10 cents to produce. The computer monopolist can rent the computer for 90 cents a second and allow the user to buy cards on the open market for 1 cent, or, if tying is permitted, he can require the user to buy cards from him at 10 cents a card—but in that case he must reduce his machine rental charge to nothing, so what has he gained?”21 This implies that the monopolist firm never has the incentive to bundle for the purpose of monopolizing the adjacent good market. As a result, price discrimination, as opposed to leverage, has come to be seen as the main motivation for tying until the theory was revived by Whinston (1990), which spawned a resurgence of interest in the theory. 3.2.2. Strategic Leverage Theory In a very influential paper, Whinston (1990) has shown that the Chicago school arguments can break down under certain circumstances. In particular, he considered a model in which the market structure in the tied good market is oligopolistic and scale economies are present, and showed that bundling can be an effective and profitable strategy to alter market structure by making continued operation unprofitable for tied good rivals. Whinston sets up a model with firms that differ in their production costs. To apply his intuition to information goods, I modify his model and present it in a setting with products of different qualities. To understand Whinston's argument, consider the following simple model. There are two independent products, A and B. (p. 288) They are unrelated in the sense that they can be consumed independently and their values to consumers are independent of whether they are consumed separately or together.22 Consumers, whose total measure is normalized to 1, are assumed to be identical and have a unit demand for each product. The market for product A is monopolized by firm 1 with unit production cost of zero. It is assumed that entry into market A is not feasible. The product A is valued at vA by consumers. The market for product B, however, can be potentially served by two firms, firm 1 and firm 2. Unit production cost for product B is also zero for both firms. However, the two firms’ products are differentiated in terms of quality. Consumers’ valuation of product B produced by firm 1 and firm 2 are respectively given by vB1 and vB2 . The game is played in the following sequence. In the first stage, the monopolistic supplier of product A (firm 1) makes a bundling decision on whether to bundle A with another product B. In the second stage, firm 2 makes an Page 10 of 23 Bundling Information Goods entry decision after observing the incumbent firm's bundling decision. The entry entails sunk fixed costs of K. If there is entry by the rival firm, a price game ensues in the final stage. The bundling decision is assumed to be irreversible. I apply backward induction to solve the game. If there is no entry, firm 1 charges (vA + vB1) if the products are bundled, and vA and vB1 for products A and B, respectively, if the two products are not bundled. In either case, firm 1 receives the monopoly profits of (vA + vB1) without entry by firm 2. Now suppose that there is entry by firm 2. The pricing stage outcome depends on the bundling decision. Suppose that the monopolist bundles product A and B and charges price P for the bundled product. In this case, consumers have two choices. The first option is to buy the bundled product from the monopolist at the price of P and the second one is to buy only product B from firm 2. For the first option to be chosen by the consumers, P should satisfy the following condition: This implies that the maximum price the tying firm can charge for the bundled product is given by P = (vA + vB1− vB2 ). Without any marginal cost, firm 1's profit is also given by (vA + vB1 − vB2 ). Of course, firm 1 will sell the bundle only if this profit is nonnegative. In other words, which firm will capture market B depends on the comparison of (vA + vB1) and vB2 . If the former is higher than the latter, firm 1 will serve market B, and otherwise firm 2 will serve market B. One way to interpret the result above is that after bundling firm 1 can compete against firm 2 as if its quality of B were (vA + vB1). Alternatively, we can also interpret this result as firm 1 behaving as if its cost of B were negative at –vA. The reason is that after bundling, firm 1 can realize the monopoly surplus of vA only in conjunction with the sale of product B. Thus, the firm is willing to sell product B up to the loss of vA. Now suppose that K〈 vB2 − vB1 〈 vA. The first inequality means that firm 2 can successfully enter the market B if there was no bundling since its quality advantage is more than the sunk cost of entry. However, the second inequality implies that the quality advantage for firm 2 is not sufficiently high to compete against the bundled products since firm 1 is still able to sell the bundled products with a positive profit (p. 289) even if firm 2 priced its product at its marginal cost zero. Thus, firm 2 is foreclosed from market B since it cannot recoup its sunk cost of entry when firm 1 engages in bundling. Essentially, bundling converts competition into an “all-out war” for firm 1 and serves as a commitment mechanism to price more aggressively whereas unbundling segments the markets and allows limited and “localized” market competition. The Chicago school critique of the leverage theory missed this “strategic effect” due to their adherence to the assumption of competitive, constant returns-to-scale structure in the tied good market. It is important to keep in mind, however, that in Whinston's basic model, inducing the exit of the rival firm is essential for the profitability of tying arrangements.23 If the bundling firm fails to induce exit of the rival firm, the bundling strategy also hurts the bundler as it intensifies competition.24 Notice that in Whinston's model, unbundling is ex post optimal once entry takes place. Thus, the monopolist should have the commitment ability to bundle to deter entry. Technical bundling through product design is one way to achieve such commitment. For instance, in the Microsoft bundling case, the Internet browser and media player programs are so tightly integrated with the rest of its OS, they cannot be removed without destabilizing the OS.25 In this regard, it is worth mentioning Peitz (2008) who provides a model of entry deterrence that does not require the commitment of the firm to bundle. More specifically, he considers a market with two products, one of which is monopolized with consumers’ reservation values uniformly distributed. In the competitive market, the products are horizontally differentiated. In such a setup, he shows that bundling is always optimal irrespective entry, and thus credible. 3.2.3. Dynamic Leverage Theory The analysis of Whinston has been subsequently extended in several directions by various authors such as Carlton and Waldman (2002) and Choi and Stefanadis (2001). Whinston's model assumes that entry into the monopolized market is impossible and shows how tying can be used to extend monopoly power in one market into an otherwise competitive market. These papers, in contrast, consider an oligopolistic environment and show that bundling can be used to deter entry into a complementary market to preserve the monopolized market or Page 11 of 23 Bundling Information Goods strengthen market power across several markets. The basic intuition for these results in Carlton and Waldman (2002) and in Choi and Stefanadis (2001) is that entry in one market is dependent on the success of entry in a complementary market. As is presented in more detail later, Carlton and Waldman develop a dynamic model where an entrant with a superior complementary product today can enter the primary product market in the future. They show that bundling can be used to deny an entrant sales when it has only the complementary product and this reduced sales today can prevent future entry into the primary market in the presence of scale economies. Choi and Stefanadis (2001) model a situation in which the incumbent firm is a monopolist in markets A and B and the two products are perfect complements. Neither has value to consumers without the other. The primary vehicle for entry is innovation, but success in innovation is uncertain. By tying the two products, the incumbent ensures that potential entrants must invest in both products and must innovate successfully in both in order to enter in the next period. Under some conditions, the tie can be profitable by discouraging investment by potential entrants, and thus reducing or even eliminating the risk of future entry. Choi (2004) develops a model that also involves current investment in innovation determining future sales. However, in contrast to the Choi and Stefanadis model, this model assumes that demand for the two goods is independent. That is, there is no relationship between consumers’ demand for the two goods. In Choi's model, tying by the monopolist initially intensifies price competition, because the monopolist must lower the price of the tied combination to attract buyers. This reduces the monopolists’ own profits in the first period. With tying, however, the monopolist has a greater incentive to invest in innovation and the other firm has a lesser incentive. Under certain conditions, tying can increase profits for the monopolist.26 3.3. The Leverage Theory of Bundling Applied to Information Goods In this section, we discuss the implications of bundling for competition in various information good industries. In particular, several information goods are characterized by network effects with demand-side scale economies. The literature distinguishes two types of network effects. The notion of direct-network effects refers to a situation in which each user's utility of adopting a particular good increases as more people adopt the same good. The most commonly cited examples of direct network effects include communications services (telephone, fax, email, etc) and languages. As these examples suggest, direct network effects arise through the ability to interact or communicate with other people. Indirect network effects typically arise through complementary products. For instance, in system markets comprised of hardware and software, consumers do not derive additional direct utilities from consumers using the same hardware. However, each consumer's utility can be increasing in the number of available software titles. As more consumers adopt a particular platform of hardware, more software titles compatible with that platform will be available. Thus, consumers indirectly benefit from the number of adopters of the same hardware. 3.3.1. Bundling of Information Goods with Direct Network Effects Certain information goods exhibit network effects in consumption. One prominent example is software. For instance, the use of the same software allows easy collaboration and sharing of files among co-workers and friends. Carlton and Waldman's (2002) model, inspired by the Microsoft antitrust case, shows that the presence of (direct) network externalities for the complementary good can result in the strategic use of bundling to deter entry into the primary market. More specifically, Carlton and Waldman's model involves an incumbent that is a monopolist in the primary market (say, operating system), called A.27 The monopolist and a potential entrant can produce a complementary product (say, web browser), B. If the entrant is successful in B in the current period, it can enter A in the future and challenge the monopolist.28 However, if the entrant is not successful in B during this period, it cannot enter market A in the future. In their model, by bundling A and B during this period, the monopolist can prevent successful entry in B this period and thus protect its existing monopoly in A against subsequent entry. 3.3.1. Bundling in Two-Sided Markets with Indirect/Inter-Group Network Effects There are many markets where indirect network effects arise through platforms that enable interactions between two distinct groups of end users. For instance, game developers and gamers interact through video game platforms such as Nintendo, Sony Play Station, and Microsoft X-Box, in which each side benefits more from the presence of 29 Page 12 of 23 Bundling Information Goods more participants on the other side.29 This type of platform competition has been analyzed under the rubric of “two-sided markets” in the literature.30 A few studies analyze the effects of bundling arrangements on platform competition in two-sided markets, partly motivated by recent antitrust cases involving Microsoft and credit card payment systems. In the European Commission Microsoft case, it has been alleged that the company's bundling practice of requiring Windows operating system users to accept its Windows Media Player software is predatory and hurts digital media rivals such as RealNetworks. In the streaming media software case, content providers and consumers constitute the two sides of the market. For instance, the more content available in streaming media, the more valuable media player programs become, and vice versa. Choi (2007) develops a model that reflects the Microsoft case in the EC.31 More specifically, the model assumes that there are two intermediaries competing for market share within each group. There is free entry in the market for content provision. Content providers are heterogeneous in their fixed cost of creating content. The choice of consumers’ platform is analyzed by adopting the Hotelling model of product differentiation in which the two platforms are located at the two extreme points of a line. Consumers are uniformly distributed along the line and each consumer's utility of participating in a platform depends on the number of content providers on the same platform. In such a model, Choi (2007) shows that bundling can be a very effective mechanism through which a dominant firm in a related market can penetrate one side of the two-sided market to gain an advantage in competition for the other side. The welfare effect of bundling is, however, ambiguous and depends on the relative magnitude of inter-group externalities and the extent of product differentiation. If the extent of inter-group externalities is (p. 292) significant compared to that of product differentiation, bundling can be welfareenhancing, as the benefit from internalizing the inter-group network externalities can outweigh the loss of product variety. Amelio and Jullien (2007) provide another analysis of tying in two-sided markets. They consider a situation in which platforms would like to set prices below zero on one side of the market to solve the demand coordination problem in two-sided markets, but are constrained to set non-negative prices. In the analysis of Amelio and Jullien, tying can serve as a mechanism to introduce implicit subsidies on one side of the market in order to solve the aforementioned coordination failure in two-sided markets. As a result, tying can raise participation on both sides and can benefit consumers in the case of monopoly platform. In a duopoly context, however, tying also has a strategic effect on competition. They show that the effects of tying on consumer surplus and social welfare depend on the extent of asymmetry in externalities between the two sides. Gao (2009) considers a new type of bundling in two-sided market, which he calls “hybrid” bundling. He observes that there are many examples of “mixed” two-sided markets in which economic agents participate in both sides of the market. For instance, people can participate in online trading both as a seller and as a buyer. Hybrid bundling is a two-part tariff system in which a bundled membership fee is charged for access to both sides of the market and two separate transaction fees that need to be paid for each transaction depending on which side the agent is making the transaction. In other words, hybrid bundling is a combination of pure bundling of membership fees for two-sides and unbundled transaction fees on each side. He provides conditions under which hybrid bundling is profitable and indicates main factors that favor it. One of the factors relevant to two-sided markets with information goods is scope of economies in providing two services to the same user compared to two different users. In the provision of information goods, the fixed cost of servicing a user usually comes from the registration process and there could be significant economies of scope if necessary information can be collected and verified in one signing-up process rather than two separate occasions. 3.3.3. Virtual Bundling In systems markets where several complementary components work together to generate valuable services, a firm can practice “virtual bundling” by making its components incompatible with components made by other firms. Even if all component products are sold separately, consumers do not have an option to mix and match components from different manufacturers and purchase the whole system from one vendor in the absence of compatibility. The possibility of virtual bundling arises if compatibility requires adoption of a common standard and consequently can be achieved only with the agreement of all firms. In the software case, for instance, virtual bundling can take place if a dominant firm has proprietary interface Page 13 of 23 Bundling Information Goods information and refuses to license it to prevent interoperability with third parties’ complementary software. Consider the recent (p. 293) antitrust cases concerning Microsoft in the US and Europe. One of the issues in the cases evolved around the interoperability information. Microsoft was alleged to deny disclosing interface information which rival work group server operating system vendors needed to interoperate with Microsoft's dominant Windows PC operating system. 32 A similar issue has arisen in an antitrust investigation against Qualcomm in Korea in which the Korean Fair Trade Commission was concerned with the possibility that nondisclosure of ADSP (Application Digital Signal Processor) interface information may restrain competition in the mobile multimedia software market.33 The non-disclosure of proprietary interoperability information constitutes virtual bundling and its strategic effects can be analyzed in a similar manner. 3.3.4. The Importance of Multihoming Most of the existing literature on strategic theory of bundling assumes single-homing, that is, consumers buy from only one vendor/product for each product category. In addition, none of these papers in the tying literature seriously take into consideration the possibility of multihoming. Carlton and Waldman (2002), for instance, assume that “if a consumer purchases a tied good consisting of one unit of the monopolist's primary good and one unit of its complementary good, then the consumer cannot add a unit of the alternative producer's complementary good to the system (italics added, p. 199).” In other words, either they do not allow the possibility of multihoming or multihoming does not arise in equilibrium.34 However, it is common in many markets that consumers engage in multihoming, that is, consumers purchase multiple products (or participate in multiple platforms). Multihoming is especially important in markets with network effects because it allows consumers to reap maximal network benefits. For instance, consider the digital media market as a two-sided market in which content providers and end users constitute each side. In this market, if more content is provided in the format of a particular company, then more users will use Media Player of such company to access such content. Moreover, if more users select a particular company's Media Player, then content providers will obviously have an incentive to provide their content in that particular format, creating indirect network externalities. However, in the digital media market, many users have more than one media player and many content providers offer content in more than one format. One implication of the single-homing assumption is that if consumers are not allowed to participate in more than one platform, tying automatically leads to monopolization on the consumer side. This in turn implies that content providers have no incentives to provide any content for the rival platform. As a result, the rival platform is foreclosed on both sides of the market. With multihoming possibilities, bundling does not necessarily imply “tipping” and market foreclosure, which can have important implications for market competition. Choi (2010a) constructs a model of two-sided markets that explicitly takes multihoming into consideration. To derive an equilibrium in which both content (p. 294) providers and consumers multihome, he assumes that there are two types of content available. One type of content is more suitable for one of the two platforms (formats) whereas the other type of content is suitable for both platforms. Alternatively, we can interpret the first type of content is due to exclusive contracts between the platforms and content providers. More specifically, the total measure of content potentially available for each format is normalized to 1. Among them, the proportion of λ is of the first type and thus can be encoded only for a particular format whereas (1 − λ) can be also encoded in the other format. The existence of exclusive content available for each format creates incentives for consumers to multihome. When the second type of content is encoded for both formats, content providers are said to multihome. The consumer side of the market is once again modeled a la Hotelling. The only modification is that consumers are now allowed to multihome. As a result, there are three choices for consumers assuming that the market is covered. Consumers can choose to either single-home or multihome. If they decide to single-home, they choose one of the two platforms to participate in. See Figure 11.2. Page 14 of 23 Bundling Information Goods Click to view larger Figure 11.2 Two-Sided Markets with Multihoming. Under such a framework, Choi (2010a) derives conditions under which “multihoming” equilibria exist under both tying and no tying and analyzes welfare implications. It is shown that the beneficial effects of tying come mainly from wider availability of exclusive content. According to the model, tying induces more consumers to multihome and makes platform-specific exclusive content available to more consumers, which is also beneficial to content providers. There are two channels for this to happen. First, tying induces all consumers to have access to exclusive content for platform A. Second, the number of consumers who have access to exclusive content for platform B also increases. However, as in the single-homing consumers case, there are less desirable matches between the consumers and platforms, leading to higher overall “transportation costs” in the product space. Thus, the overall welfare effects can be in general ambiguous. However, the simple structure of the model posited in Choi (2010a) yields an unambiguous answer that tying is welfare-enhancing. (p. 295) To explore the role of multihoming in the model, he also considers a hypothetical situation in which bundling prevents consumers from multihoming, thus leading to the foreclosure of competing products. For instance, the monopolist engages in technical tying in which it designs its product in such a way that a competitor's product cannot interoperate with the tying product.35 Without multihoming, all consumers will use the tied product only in the two-sided market. This implies that all content is provided for the bundling firm's format and that exclusive content for the rival format will be no longer available. In such a case, it can be shown that tying is unambiguously welfare-reducing, which is in sharp contrast to the result obtained with the assumption of multihoming. The upshot of the model in Choi (2010a) is that we derive diametrically opposite results depending on whether or not we allow the possibility of multihoming after tying. This simple result undoubtedly comes from many special features of the model. Nonetheless, the model highlights the importance of explicitly considering the role of multihoming in the antitrust analysis of two-sided markets and provides caution in simply taking the theoretical results derived in models with “single-homing” and extrapolating to markets where “multihoming” is prevalent. Multihoming has potential to counteract the tendency towards tipping and the lock-in effects in industries with network effects. As a result, bundling does not necessarily lead to the monopolization of the market with multihoming. Considering the prevalence of multihoming and exclusive content in information goods industries, it would be worthwhile to enrich Choi's model. First, the model assumes that there is an exogenous amount of exclusive content available for each format. It would be desirable to have a model where exclusivity is endogenously created through the use of exclusive contracts.36 Another avenue of research would be to explore the implications of bundling and multihoming for the incentives to innovate and create new content in information goods industries. 3.3.5. Bundling as a Rent Extraction Device with Multihoming When multihoming is possible, consumers are not obligated to use the products included in the bundle and may opt to use alternative products from other vendors. Carlton, Gans, and Waldman (2010) provide a novel model that explains why a firm would “tie a product consumers do not use.” In their model, there is a monopolist of a primary product and a complementary product that can be produced both by the monopolist and an alternative producer.37 In this setting of complementary products, the monopolist's tied product provides the consumer with a backup option. The presence of that option affects consumer willingness to pay for the rival's complementary product, which in turn affects pricing of the monopolized product due to the complementary nature of the two Page 15 of 23 Bundling Information Goods products. In this model, the monopolist's tied product is never used by consumers. Nonetheless, tying can serve as a rent-extracting mechanism, and is thus profitable.38 The practice is obviously inefficient to the extent that it entails additional costs to include the product not used. However, the practice is not exclusionary and does not foreclose the market (p. 296) as in models of strategic tying with single-homing such as Whinston (1990), Choi and Stefanadis (2001), and Carlton and Waldman (2002). 3.3.6. Implications of Bundling for Merger Analysis The possibility of bundling strategy also has important implications for merger analysis. When a merger takes place among firms that produce different products, the merged entity can now afford to offer bundles that were not feasible when the firms remained as separate firms. Choi (2008) analyzes the effects of mergers in complementary markets when the merged firm can engage in bundling. More specifically, he considers two complementary components, A and B, which consumers combine in fixed proportions on a one-to-one basis to form a final product. For instance, A and B can be considered as operating systems and application software, respectively, to form a computer system. He assumes that there are two differentiated brands of each of the two components A (A1 and A2 ) and B (B1 and B2 ). This implies that there are four system goods available, A1B1, A1B2 , A2 B1, and A2 B2 . In such a framework, Choi (2008) analyzes how the market equilibrium changes after a merger between A1 and B1 when the merged firms engage in mixed bundling. He shows that the merged firm's bundle price is reduced compared to the sum of individual prices before the merger, as the merged firm internalizes the pricing externalities arising from the complementarity of the two component products. At the same time, the merged entity raises the prices of its stand-alone components, relative to their levels prior to the merger. In response to the price cut by the merged firm for their bundled system and the price increase for the “mix-and-match” systems, the independent rivals cut price in order to retain their market shares. However, their price cut is less than the one reflected in the bundle. In conjunction with higher component prices by the merged firm, independent firms lose market shares compared to the premerger situation. As a result, the merging firms’ market share increases at the expense of the independent firms. The independent firms unambiguously suffer from the combination of a loss of market share and the need to cut prices. The overall welfare effect is ambiguous because a merger with bundling in complementary markets may have both anticompetitive effects and efficiency benefits. The efficiency benefits, for instance, take the form of internalizing pricing externalities for the merged firm. Consumers as a group may suffer. With heterogeneous consumer preferences, some buyers gain and others lose. For instance, those who previously purchased both products from the two merging firms would gain due to the lower bundle price. However, those who purchased a “mix and match” system and wished to continue doing so would suffer due to the increased stand-alone prices charged by the merged firm. It is possible that overall consumer surplus may decline. In addition, the potential anticompetitive effects may take the form of market foreclosure if the financial impact of the merged firm's bundling made its rivals unable to cover their fixed costs of production. Alternatively, adverse welfare impacts may also arise from changed R&D investment incentives. (p. 297) Choi (2008) also analyzes the effects of a merger with pure bundling under which the firm only sells the bundle and does not make the individual components available separately. Consistent with the results in Whinston (1990), he shows that pure bundling is not a profitable strategy if it fails to induce exit of rival firms because it intensifies competition. However, as in Whinston (1990), pure bundling can still be profitable if the exclusion of rivals through predation is possible with pure bundling, but not with mixed bundling.39 3.4. Case Studies I provide a short discussion of two recent bundling cases in information good industries and their welfare implications. 3.4.1. The Microsoft Antitrust Cases Microsoft's bundling practices have encountered antitrust investigations on both sides of the Atlantic. In the European antitrust case, it has been alleged that the company's tying practice of requiring Windows operating Page 16 of 23 Bundling Information Goods system (OS) users to accept its Windows Media Player software is anticompetitive and hurts digital media rivals such as RealNetworks. The European Commission (EC) also alleged that Microsoft had leveraged its market power in the PC OS market into the adjacent work group server OS market by withholding interoperability information. Even though the interoperability issue was couched in terms of information disclosure, the economic effect of restricting interoperability between Microsoft's PC OS and rival work group server operating systems is akin to bundling of Microsoft's PC OS and its own work group server OS because incompatibility between complementary components from different manufacturers deprives consumers of the ability to mix-and-match and constitutes a virtual bundling, as we discussed earlier. On March 24, 2004, the EC ruled that Microsoft is guilty of abusing the “near-monopoly” of its Windows PC operating system and fined it a record 497 million Euros ($613 million). As a remedy, the EC required Microsoft to make available a version of the Windows OS that either excludes the company's Media Player software or includes competing products, and to disclose “complete and accurate specifications for the protocols used by Windows work group servers” to achieve full interoperability with Windows PCs and servers.40 The ruling was appealed, but upheld by the Court of First Instance on September 17, 2007.41 In the U.S., the Department of Justice (DOJ) alleged that Microsoft engaged in a variety of exclusionary practices. In particular, Microsoft's practice of bundling its web browser with the dominant Windows PC OS was alleged to be anticompetitive along with other exclusive contracts. The case was eventually settled with a consent decree that mandated disclosure of application programming interfaces (API) information and required Microsoft to allow end users and OEMs to enable or remove access to certain Windows components or functionalities (such as Internet (p. 298) browsers and media players) and designate a competing product to be invoked in place of Microsoft software. However, the DOJ did not prevent Microsoft from bundling other software with Windows OS in the future. It is a difficult question to answer whether the DOJ in the US or the EC made the right decisions concerning Microsoft's bundling practices.42 The welfare implications of bundling arrangements are in general ambiguous because bundling could have efficiency effects even when it has harmful exclusionary effects. The literature suggests that bundling can be exclusionary. At the same time, however, there may be offsetting effects of bundling such as enhanced performance due to a seamless integration of products and reduced costs of distribution if the bundled goods are often used together. The appropriate antitrust policy concerning bundling will be ultimately an empirical question that assesses possible efficiency effects against potential anti-competitive effects, and depend on the specifics of the case. 3.4.2. Bundling in the Publishing Industry One of the major developments in the publishing industry is the distribution of content in digital forms through the Internet. For instance, site licensing of electronic journals revolutionized the way academic articles are accessed. For academic journals, major publishers also engage in bundling schemes in which individual journal subscriptions may not be cancelled in their electronic format. The leading example is the ScienceDirect package by Elsevier with access to 2,500 journals and more than nine million full-text articles (as of July 2010). Jeon and Menicucci (2006) provide an analysis of publishers’ incentives to practice (pure) bundling and its effects on social welfare, and derive implications for merger analysis.43 They assume that each library has a fixed budget that can be allocated between journals and books. In addition, they assume that the library's utility from purchasing books is a concave function of its expenditure on books whereas the value of each journal is independent of whether or not the library purchases any other journals. In their model, the effects of bundling on pricing thus arise solely through its impact on the library's budget allocation between journals and books. More specifically, they show that bundling entails two distinct effects. First, it has the direct effect of softening competition from books. To understand this, consider a publisher with two journals of the same value v for a library. With independent pricing, the publisher charges the same price p for them. Now suppose that the publisher bundles the two journals and charges the price of 2p for the bundle. Then, the library is strictly better off to buy the bundle than alternatively spending 2p on books due to the diminishing marginal utility of spending money on books. As a result, the publisher can charge more than 2p for the bundle and still induce the library to buy the bundle. Note that this effect increases with the size of the bundle. Second, bundling strategy has an indirect effect of negative externalities on other publishers. The reason is that a higher price for the bundle implies less money left for other publishers and books. As a result, other publishers need to cut the price of their journals in competition with books. Bundling is thus a profitable (p. 299) and credible strategy in that it increases the bundling firm's profit and may make the rival firms unable to sell their journals. However, they show that bundling is socially welfare-reducing. They also Page 17 of 23 Bundling Information Goods explore implications of mergers and show that any merger among publishers is profitable but reduces social welfare. Their analysis thus has important implications for recent consolidations in the academic journal publishing industry when the possibility of bundling is considered. 4. Conclusions This chapter has conducted a highly selective review of the literature on bundling and discussed how the theory can be applied to information goods. I focused on two stands of the bundling literature: one on price discrimination and the other on strategic leverage theory. From the perspective of price discrimination, recent advances in the Internet technology present both opportunities and challenges for bundling strategies. With the digitalization of information goods and the Internet as a new distribution channel, there are opportunities for a large scale bundle as a price discriminating instrument because there is less concern for the inefficiency that may arise by forcing consumers to purchase goods that they value below marginal cost, which is very likely with physical goods. At the same time, the Internet technology enables firms to collect information about customers’ purchase behavior and use sophisticated data-mining software tools to infer customers’ preferences. The Internet also allows for sophisticated mass customization that produces goods tailored to each customer's needs without significantly sacrificing cost efficiency. This development may obviate the need to use bundling as a price discrimination device. In this new environment, firms are constantly experimenting to figure out the best ways to capture consumer surplus, and the best business models are still in flux with ever evolving technologies available. However, consumers’ strategic responses to such individualized pricing may make such strategies less profitable, and bundling is expected to remain an important pricing strategy for information goods. The leverage theory of bundling has important implications for antitrust analysis. Bundling typically entails efficiency effects such as a seamless integration of products and reduced distribution costs even when it has harmful exclusionary effects. As such, there seems to be a consensus among economists that bundling should not be treated as per se violation of antitrust laws and that the rule of reason should be adopted in the assessment of tying arrangements. It would be an important task to come up with general guidelines to assess possible efficiency effects of bundling against potential exclusionary effects. In particular, tying can be a very effective mechanism through which a dominant firm in a related market can penetrate one side of the two-sided market to gain an advantage in competition for the other side. As such, we are expected to observe more tying cases in two-sided (p. 300) markets, and it is essential to understand the impacts of tying on competition in such markets and their welfare consequences. In this regard, one important element to consider is whether multihoming is a viable option for relevant players. References Acquisti, A., Varian, H.R., 2005. Conditioning Prices on Purchase History. Marketing Science 24(3), pp.1–15. Adams, W.J., Yellen, J.L., 1976. Commodity Bundling and the Burden of Monopoly. Quarterly Journal of Economics 90(3), pp. 475–498. Amelio, A., Jullien, B., 2007. Tying and Freebie in Two-Sided Markets, IDEI Working Paper No. 445. Armstrong, M., 1999. Price Discrimination by a Many-Product Firm. Review of Economic Studies 66(1), Special Issue, pp. 151–168. (p. 303) Armstrong, M., 2006. Competition in Two-Sided Markets. RAND Journal of Economics, pp. 668–691. Armstrong, M., 2011. Bundling Revisited: Substitute Products and Inter-Firm Discounts. Economics Series Working Papers 574, University of Oxford. Arrow, K., 1962. Economic Welfare and the Allocation of Resources for Invention. The Rate and Direction of Inventive Activity: Economic and Social Factors, pp. 609–626. Page 18 of 23 Bundling Information Goods Bakos, Y., Brynjolfsson, E., 1999. Bundling Information Goods: Pricing, Profits and Efficiency. Management Science 45(12), pp.1613–1630. Bakos, Y., Brynjolfsson, E., 2000. Bundling and Competition on the Internet. Marketing Science 19(1), pp.63–82. Belleflamme, P., Peitz, M., 2010. Industrial Organization: Markets and Strategies. Cambridge, U.K.: Cambridge University Press. Bork, R.H., 1978. The Antitrust Paradox: A Policy at War with Itself. New York, Basic Books. Bowman, W., 1957. Tying Arrangements and the Leverage Problem, Yale Law Journal 67, pp.19–36. Bulow, J.I., Geanakoplos, J.D, Klemperer, P.D., 1985. Multimarket Oligopoly: Strategic Substitutes and Complements. Journal of Political Economy 93, 488–511. Carbajo, J., De Meza, D., Seidman, D.J., 1990. A Strategic Motivation for Commodity Bundling, Journal of Industrial Economics 38, pp. 283–298. Carlton, D.W., Gans, J., Waldman, M., 2010. Why Tie a Product Consumers Do Not Use? American Economic Journal: Microeconomics 2, pp. 85–105. Carlton, D.W., Waldman, M., 2002. The Strategic Use of Tying to Preserve and Create Market Power in Evolving Industries. RAND Journal of Economics, pp. 194–220. Chen, Y., 1997. Equilibrium Product Bundling. Journal of Business 70, pp. 85–103. Chen, Y., Riordan, M., 2010. Preference Dependence and Product Bundling. Unpublished manuscript. Choi, J.P., 1996. Preemptive R&D, Rent Dissipation, and the Leverage Theory. Quarterly Journal of Economics, pp. 1153–1181. Choi, J.P., 2004. Tying and Innovation: A Dynamic Analysis of Tying Arrangements. The Economic Journal 114, pp. 83–101. Choi, J.P., 2007. Tying in Two-Sided Markets with Multi-Homing. CESifo Working Paper No. 2073. Choi, J.P., 2008. Mergers with Bundling in Complementary Markets. Journal of Industrial Economics, pp. 553–577. Choi, J.P., 2010a. Tying in Two-Sided Markets with Multi-Homing. Journal of Industrial Economics 58, pp. 560–579. Choi, J.P., 2010b. Compulsory Licensing as an Antitrust Remedy. The WIPO Journal 2, pp. 74–81. Choi, J.P., Lee, G., Stefanadis, C., 2003. The Effects of Integration on R&D Incentives in Systems Markets. Netnomics, Special Issue on the Microeconomics of the New Economy, pp. 21–32. Choi, J.P., Stefanadis, C., 2001. Tying, Investment, and the Dynamic Leverage Theory. RAND Journal of Economics 32, pp. 52–71. Chu, C.S., Leslie, P., Sorensen, A.T., 2011. Bundle-Size Pricing as an Approximation to Mixed Bundling. American Economic Review 101, pp. 263–303. (p. 304) Chuang, J.C.-I., Sirbu, M.A., 1999. Optimal Bundling Strategy for Digital Information Goods: Network Delivery of Articles and Subscriptions. Information Economics and Policy 11(2), pp. 147–176. Doganoglu, T., Wright, J., 2010. Exclusive Dealing with Network Effects. International Journal of Industrial Organization 28, pp. 145–154. Farrell, J., Katz, M. L., 2000. Innovation, Rent Extraction, and Integration in Systems Markets. Journal of Industrial Economics 48, pp. 413–432. Flores-Fillol, R., Moner-Colonques, R., 2011. Endogenous Mergers of Complements with Mixed Bundling. Review of Industrial Organization 39, pp. 231–251. Page 19 of 23 Bundling Information Goods Fudenberg, D., Tirole, J., 1984. The Fat Cat Effect, the Puppy Dog Ploy and the Lean and Hungry Look, American Economic Review 74, pp. 361–368. Fudenberg, D., Villas-Boas, J. M., 2006. Behavior-Based Price Discrimination and Customer Recognition. In: T.J. Hendershott (Ed.), Economics and Information Systems: Handbooks in Information Systems Vol. 1, Amsterdam, Elsevier, pp. 377–436. Gans, J., King, S., 2006. Paying for Loyalty: Product Bundling in Oligopoly. Journal of Industrial Economics 54, pp. 43–62. Gao, M., 2009. When to Allow Buyers to Sell? Bundling in Mixed Two-Sided Markets. Unpublished manuscript. Available at: http://phd.london.edu/mgao/assets/documents/Ming_Gao_JM_Paper.pdf. Jeon, D.-S., Menicucci, D., 2006. Bundling Electronic Journals and Competition among Publishers. Journal of the European Economic Association 4(5), pp. 1038–1083. Kenney, R.W., Klein, B., 1983. The Economics of Block Booking. Journal of Law and Economics 26(3), pp. 497–540. Loginova, O., Wang, X. H., 2011. Customization with Vertically Differentiated Products. Journal of Economics and Management Strategy 20(2), pp. 475–515. Matutes, C., Regibeau, P., 1988. Mix and Match: Product Compatibility without Network Externalities. RAND Journal of Economics 19(2), pp. 221–234. McAfee, R.P., McMillan, J., Whinston, M.D. 1989. Multiproduct Monopoly, Commodity Bundling, and Correlation of Values. Quarterly Journal of Economics 114, pp. 371–384. Mialon, S.H., 2011. Product Bundling and Incentives for Merger and Strategic Alliance. Unpublished manuscript. Nalebuff, B., 2000. Competing Against Bundles. In: P. Hammond, Myles, G. (Eds.), Incentives, Organization, and Public Economics, Oxford, Oxford University Press, pp. 323–336. Nalebuff, B., 2004. Bundling as an Entry Barrier. Quarterly Journal of Economics 119, pp. 159–188. Peitz, M., 2008. Bundling May Blockade Entry. International Journal of Industrial Organization 26(1), pp. 41–58. Posner, R.A., 1976. Antitrust Law: An Economic Perspective, Chicago: University of Chicago Press. Rochet, J.-C., Tirole, J., 2006. Two-Sided Markets: A Progress Report. RAND Journal of Economics 37(3), pp. 645– 667. Rochet, J.-C., Tirole, J., 2008. Tying in Two-Sided Markets and the Honor All Cards Rule. International Journal of Industrial Organization 26(6), pp. 1333–1347. Schmalensee, R.L., 1984. Gaussian Demand and Commodity Bundling. Journal of Business 57, pp. 211–230. (p. 305) Shiller, B., Waldfogel, J., 2011. Music for a Song: An Empirical Look at Uniform Song Pricing and Its Alternatives. Journal of Industrial Economics 59, pp. 630–660. Simon, H.A., 1971. Designing Organizations for an Information-Rich World. In: M. Greenberger (Ed.), Computers, Communication, and the Public Interest, Baltimore, The Johns Hopkins Press, pp. 37–72. Stigler, G.J. 1963. United States v. Loew’s, Inc.: A Note on Block Booking. Supreme Court Review 1963, pp. 152– 157. Taylor, C.R., 2004. Consumer Privacy and the Market for Customer Information. RAND Journal of Economics 35(4), pp.631–650. Trachtenberg, J.A. 2011. Sellers of E-Books Bundling Titles to Promote Authors, New and Old. Wall Street Journal, Feb 11, 2011, p. B6. Page 20 of 23 Bundling Information Goods Whinston, M.D. 1990. Tying, Foreclosure, and Exclusion. American Economic Review 80, pp. 837–859. Whinston, M.D., 2001. Exclusivity and Tying in U.S. v. Microsoft: What We Know, and Don’t Know. Journal of Economic Perspectives 15, pp. 63–80. Yoo, C.S., 2009. The Convergence of Broadcasting and Telephony: Legal and Regulatory Implications. Communications and Convergence Review 1(1), pp. 44–55. Notes: (1.) At Random House, digital titles already account for nearly 50 percent of revenue for some fiction best sellers. See Trachtenberg (2011). (2.) Bundling can also be practiced by a third party. For instance, online travel agencies sell vacation packages that combine flight, hotel, and rental car with substantial savings from what would cost if they are purchased separately. The package prices mask individual component prices. The hotels and airlines are more willing to give lower prices not available otherwise if they do not have to show their individual prices. (3.) Microsoft Office Suite, for instance, includes word processor (Word), spreadsheet (Excel), data management (Access), Email (Outlook), and presentation (PowerPoint) programs. (4.) A variation on this theme is the metering argument in which the purchase of an indivisible machine is accompanied by the requirement that all complementary variable inputs be purchased from the same company. By marking-up the variable inputs above marginal cost, the seller can price discriminate against intense users of the machine with the sale of variable inputs as a metering or monitoring device for the intensity of the machine usage. (5.) Independent pricing is a special case of mixed bundling in which the bundle price is infinity and pure bundling is a special case of mixed bundling in which component prices are set at infinity. (6.) A copula is a multivariate uniform distribution that couples marginal distributions to form joint distributions. Sklar's Theorem states that the joint distribution for n variables can be represented by a copula and the marginal distributions or respective variables. (7.) For the exact conditions on the extent of limited positive dependence, see Chen and Riordan (2010). (8.) See also Armstrong (1999) who provides a more general but similar asymptotic result. He shows that a two-part tariff in which consumers pay a fixed fee and unit price of any product equal to its marginal cost, achieves approximately the same profit as perfect price discrimination if the number of products approaches infinity. (9.) Bakos and Brynjolfsson (1999) consider a more general case in which the valuations of each good can depend on the number of goods purchased (and thus on the bundle size) to allow the possibility that the goods in the bundle can be complements and substitutes. (10.) Adapted from Figure 1 in Bakos and Brynjolfsson (1999). (11.) See Simon (1971). (12.) See Loginova and Wang (2011) for an analysis of implications of mass customization for competition. (13.) For an excellent survey of the literature on behavior-based price discrimination, see Fudenberg and VillasBoas (2006). (14.) See Taylor (2004) and Acquisti and Varian (2005). (15.) The fully optimal mixed bundling improves the profit over individual pricing by only about 2.4 percent. However, in the presence of large fixed costs as would be typical for information goods, this difference can make a huge difference in terms of resource allocations if without bundling the firm cannot break even whereas it doe with bundling. Page 21 of 23 Bundling Information Goods (16.) Nalebuff (2000) also shows that a firm that sells a bundle of complementary products has a substantial advantage over rivals who sell component products individually. (17.) With only one second good that can be bundled, only one firm bundles in equilibrium; if both firms bundle, once again, the bundled price is driven down to the marginal costs of the bundled product because there is no longer product differentiation. With more than one second good, it is possible that both firms choose different bundles in equilibrium. (18.) See Belleflamme and Peitz (2010) for a simple and elegant exposition of Chen's model. (19.) Carbajo et al. (1990) also consider the Cournot case in the secondary market and show that bundling may not be profitable. Once again, bundling induces a favorable response from the rival, but for a different reason because prices are strategic complements while quantities are strategic substitutes. See Bulow, Bulow, Geanakoplos, and Klemperer (1985) and Fudenberg and Tirole (1984) for more details. (20.) See Gans and King (2006) who also investigate joint purchase discounts implemented by two separate firms. They consider a model in which there are two products with each product being produced by two differentiated firms. They show that when a bundle discount is offered for joint purchase of otherwise independent products, these products are converted into complements from the perspective of consumers’ purchase decisions. (21.) Richard A. Posner, “Antitrust Law: An Economic Perspective, Chicago” University of Chicago Press, 1976, p. 173. From the consumer's perspective, all that matters is the total price of data processing, not how it is divided between the two products or services, which are machine time and punch cards in this example. We know that profits are maximized when the total price is $1. Thus, any increase in price of the tied product price must be matched by a reduction in the price of the tying product. (22.) See below for the case of complementary products. (23.) See Nalebuff (2004) who also analyzes the strategy of bundling as an entry deterrence device. He considers a situation in which an incumbent with two independent products faces one product entrant, but does not know in which market the entry will take place. He shows that bundling allows the incumbent to credibly defend both products without having to lower prices in each market. Note, however, that he assumes a sequential pricing game (Stackelberg game) to derive his conclusions. If a simultaneous pricing game is assumed, the results can be markedly different. (24.) In the terminology of Fudenberg and Tirole (1984), bundling is a “top dog” strategy, while non-bundling softens price competition and is a “puppy dog” strategy. See also Bulow, Geanakoplos, and Klemperer (1985). (25.) The Microsoft bundling cases are further discussed below. (26.) See also Choi (1996). (27.) See Whinston (2001) for a more detailed discussion of the Microsoft case. (28.) Carlton and Waldman also present several variants of their model, with slightly different mechanisms for the link between the first and second period. (29.) See Chapter 4 on video game platforms by Robin Lee in this Handbook for more details. (30.) See Armstrong (2006), Rochet and Tirole (2006), and Chapters 14, 7, and 3 by Anderson, Jullien, and Hagiu, respectively, in this Handbook for an analysis of competition in two-sided markets and additional examples of twosided markets. (31.) For an analysis of the bundling practice initiated by payments card associations Visa and MasterCard, see Rochet and Tirole (2008). In this case, merchants who accept their credit cards were forced also to accept their debit cards under the so-called “honor-all-cards” rule. (32.) The Microsoft case is further discussed in Section 4. (33.) On 13 December 2010, the Korea Fair Trade Commission (KFTC) announced that Qualcomm would disclose Page 22 of 23 Bundling Information Goods ADSP interface information to third party South Korean companies to allow such companies to develop mobile multimedia software for its modem chip. The KFTC also stated that the disclosure by Qualcomm would begin within two to 10 months from the date of the announcement. See the KFTC Press Release available at http://eng.ftc.go.kr/ (34.) One notable exception is Peitz (2008) who not only allows for multihoming, but shows that multihoming actually occurs in equilibrium, that is, some consumers buy the bundle and the competitor's stand-alone product. (35.) Choi's (2010a) model suggests that such a practice can be anti-competitive. (36.) See Doganoglu and Wright (2010) for an analysis of exclusive contracts as an instrument of entry deterrence in a market with network effects and multihoming. (37.) Carlton, Gans, and Waldman (2010) say that ties are reversible when multihoming is allowed. (38.) The rent-extraction mechanism in Carlton, Gans and Waldman (2010) is similar to that in Farrell and Katz (2000) and Choi, Lee, and Stefanadis (2003). (39.) In a similar vein, Mialon (2011) builds a model in which exclusionary bundling motivates mergers. In her model, a merger is never profitable if not combined with pure bundling that leads to market foreclosure. A recent paper by Flores-Fillol and Moner-Colonques (2011) extend Choi's model to allow for both joint and separate consumption of component products. (40.) Case COMP/C-3/37.792, Microsoft v. Commission of the European Communities Decision, para 999, available at http://ec.europa.eu/comm/competition/antitrust/cases/decisions/37792/en.pdf. See Choi (2010b) for a discussion of compulsory licensing as an antitrust remedy. (41.) As of this writing, the EC is also investigation IBM for the abuse of its market dominant position in the mainframe computer market through the bundling of its hardware and mainframe OS and refusal to license information for interoperability with other software. (42.) See Whinston (2001) for an excellent discussion of the US Microsoft case. (43.) See Chuang and Sirbu (1999) for the bundling strategy of academic journals from the perspective of price discrimination. Jay Pil Choi Jay Pil Choi is Scientia Professor in the School of Economics at the Australian School of Business, University of New South Wales. Page 23 of 23 Internet Auctions Oxford Handbooks Online Internet Auctions Ben Greiner, Axel Ockenfels, and Abdolkarim Sadrieh The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0012 Abstract and Keywords This article outlines the theoretical, empirical, and experimental contributions that address, in particular, bidding behavior in Internet auctions and the auction design in single-unit Internet auctions. Revenue equivalence generally breaks down with bidder asymmetry and the interdependence of values. There is evidence on the Internet for a specific kind of overbidding, occurring when there is also competition on the supply side. There is strong evidence both for the existence of incremental and of late bidding in Internet auctions. The phenomena appear to be driven partly by the interaction of naive and sophisticated bidding strategies. Multi-unit demand typically increases the strategic and computational complexities and often results in market power problems. The emergence of Internet-specific auction formats, such as advertisement position auctions, proxy bidding, and penny auctions, initiated a whole new research literature that already feeds back into the development and design of these auctions. Keywords: Internet auctions, competition, market power, advertisement position auctions, proxy bidding, penny auctions Auctions are one of the oldest and most basic market mechanisms for selling almost any kind of goods.1 One major advantage of auctions is their ability to identify and contract the bidder(s) with the highest willingness to pay in environments with demand uncertainty (Milgrom and Weber, 1982). Traditionally auctions involved high transaction costs, because bidders typically had to gather at the same place and time. Since the high transaction costs could only be offset by even greater efficiency gains, auctions were usually only employed in situations in which price discovery was particularly pertinent. The emergence of electronic communication and the Internet lowered the transaction costs of auctions dramatically.2 Internet auctions can be held at any place and time, with bidders participating from all over the world and using proxy agents to place bids whenever necessary.3 Lower bidding costs imply that the number of participating bidders is increased, making (online) auctions even more attractive for sellers. The technological advancement gave rise to the emergence of large auction platforms such as eBay that complement their (different types of) auction services with related services such as payment handling and buyer insurance. At the same time, online auction platforms generate and record an enormous amount of detailed market data, much more both in scale and scope than available from traditional markets. This transparency has created new opportunities for research, which in turn drives innovations in market design. In fact, the academic literature on Internet auctions has expanded vastly over the last years.4 In this chapter, we provide a selective review of this research, concentrating on those aspects of auctions that are specific to or that were born out of Internet auctions.5 (p. 307) In the next section we briefly review basic auction theory. Section 3 takes a look at Internet auctions from a bidder's perspective, investigating bid amounts and bid timing. In Section 4 we consider the seller's perspective with respect to shill bidding, reserve prices, buy-now options, and the use of all-pay auctions. Section 5 discusses typical multiunit auctions conducted over the Internet. Finally, section 6 concludes with a summary and Page 1 of 27 Internet Auctions outlook for potential future research. 1. Some Basic Auction Theory In this section we briefly describe some of the most basic theoretical results on auctions that have proven useful when studying Internet marketplaces.6 We start by describing the standard types of single-unit auctions, and consider the more complex multi-unit auctions later in this chapter. Single-unit auctions can be classified into openbid and sealedbid auctions. In open-bid auctions, tentative prices are announced, and bidders indicate whether they would buy at the current price or not. Such open-bid auctions can be further classified into auctions with an ascending or descending price. In the descending-price open-bid auction (“Dutch auction”)7 , the price starts high and is lowered by the auctioneer step by step. The item is allocated to the bidder who first accepts the current price. In the ascending-price open-bid auction the price clock starts low and is then driven either by the auctioneer (“Japanese auction”) or by the bidders’ own bids (“English auction”). The auction ends once only one bidder remains in the race. The remaining bidder receives the item and pays the price at which the last bidder dropped out. In sealed-bid auctions bidders submit their bids simultaneously. Such auctions may differ in their pricing rule, with the most prominent rules being the first-price and the second-price auction. In both cases, the highest bidder wins the item, paying the own bid in the former, but the second highest bid in the latter case. Two standard models are used to describe how bidders value an item. In the private value model, each bidder knows his own value for the item (i.e., the own maximum willingness to pay). Bidders, however, do not know the exact values of the other bidders or of the seller. They only know the distributions from which each of the values is drawn and the fact that all values are drawn independently. Under the common value assumption, there is a single true value of the object that is unknown but the same for all bidders. Each bidder receives a noisy signal on the true value of the item and knows the distributions from which the own signal and the signals of the other bidders are drawn. In hybrid models, private values are either affiliated to each other or combined with a common value.8 In the following, we concentrate on the private values case, especially on the symmetric case, in which all private values are drawn from the same distribution.9 In the ascending-price open-bid auction, it is a weakly dominant strategy for each bidder to drop out of the auction when the price reaches one's own value; (p. 308) bidding more might yield a lower payoff in some cases, but never a higher payoff. Dropping out as soon as the own value is surpassed is dominant, since at any lower price a profitable opportunity may be lost, while at any higher price a loss may be incurred. As a result, the bidder with the highest value receives the good and the auction outcome is efficient. The price to be paid equals the second highest value among all bidders (possibly plus a small increment), as this is the price where the winner's strongest competitor drops out. A similar result is obtained for the second-price sealed-bid auction. Here, by auction rule, the bidder with the highest bid wins and pays the second highest bid. Thus, one's own bid only affects the probability of winning, but not the price to be paid. By bidding their own values, bidders maximize their chances of winning, but make sure that they never pay a price greater than one's own value. Thus, as in the ascending-price auction, bidding one's own value is a weakly dominant strategy in the second-price sealed-bid auction (as first observed by Vickrey, 1961). When played by all bidders, the strategy results in an equilibrium in which the highest bidder wins the auction and pays a price equal to the strongest competitor's value. Bidding in the descending-price open-bid and the first-price sealed-bid auction formats is more complex, because the winner pays a price equal to one's own bid, not knowing the bid of the strongest competitor. Thus, when submitting a bid in these auctions, a bidder has to trade-off the probability to win (which is lower with a lower bid) and the expected profit in case of winning (which is higher with a lower bid). In particular, bidders can only obtain positive payoffs if they “shade” their bids, i.e., bid less than their value. If we assume risk-neutral bidders then each bidder's equilibrium strategy is to bid the expected value of the strongest competitor conditional on having the highest value. The bidder with the highest bid then wins the auction, and the ex-ante expected auction price is equal to the expected second highest value in the bidder population. Summing up, for the case of symmetric private values (identically distributed and independently drawn), ex ante all four auction formats are efficient and yield the same expected revenue. This is the celebrated revenue equivalence theorem that was first stated by Vickrey (1961) and later generalized by Myerson (1981) and Riley 10 Page 2 of 27 Internet Auctions and Samuelson (1981) to the case of any “standard auction.”10 Revenue equivalence generally breaks down with bidder asymmetry and the interdependence of values. In the case of asymmetry, there is no clear theoretical prediction as to which type auction yields the highest revenue (Krishna, 2002). In the case of symmetric value interdependence, the linkage principle (Milgrom and Weber, 1982) generally predicts that the more information on the true value is made available by an auction mechanism, the higher the revenues. Hence, the English auction, which allows bidders to observe each others’ exit strategies, ranks first among the four classical auctions, followed by the second-price sealed-bid auction that uses the top two values to determine the price. The first-price sealed-bid auction and the descending price auction jointly come in last, due to the fact that neither procedure discharges information on bidders’ signals until after the auction is terminated. (p. 309) By setting a reserve price, sellers can increase revenues beyond the expected second highest bidder value, if they are willing to sacrifice allocational efficiency (Myerson, 1981, Riley and Samuelson, 1981). In Internet auctions, this reserve price can take the form of an open start price (below which bids are not allowed), a secret reserve (which is or is not revealed once it is reached during the course of the auction), or shill bids (i.e., the seller placing bids in his own auction). A revenue maximizing reserve price is chosen such that, in expectation, the seller can extract some of the winning bidder's surplus, thereby increasing one's own expected revenue. The strategy is successful if the reserve price is below the winner's value and above the value of the winner's strongest competitor. However, the optimal reserve price comes at the cost of lowered expected efficiency: if the seller sets the reserve price too high—higher than the highest value—then the item will not be sold. Although the theory of simple auctions yields useful orientation, it typically does not satisfactorily describe Internet auctions. In fact, in our survey we deal with a number of auction features that concern institutional and behavioral complications and are not captured by the simple theory. For instance, the models do not account for endogenous entry, flexible dynamic bidding, minimum increments, time limits, and various specific pricing and information rules. If, say, entry into an auction is endogenous, and bidders incur small but positive costs of participation or bidding, then bidders confronted with a reserve price might not enter so as to avoid potential losses. As a result, with endogenous entry and bidding costs the optimal reserve price converges to the seller's reservation value with the number of potential bidders (Levin and Smith, 1996; McAfee and McMillan, 1987; Samuelson, 1985). Moreover, the behavioral assumptions underlying the simple equilibrium analysis are often too strong. For instance, bidders might need to “construct” values, and preference construction is subject to endowment-, framing- and other cognitive and motivational effects. Also, risk-averse bidders should bid more aggressively in first-price sealed-bid and decreasing-price open-bid auctions, which thereby yield higher revenues than auctions formats with (weakly) dominant equilibrium strategies.11 Next we discuss some selected institutional and behavioral complexities that seem relevant for Internet auctions. 2. Bidding Behavior Irrespective of a broad spectrum of Internet auction designs, we recognize the basic prototypes and their strategic properties in many Internet auction markets. One popular format on auction markets is what might be interpreted as a hybrid of the English and the second-price sealed-bid auction as prominently employed by eBay (see also Lucking-Reiley, 2000a, 2000b). In the format called “proxy bidding” (sometimes also referred to as “maximum bid” or “auto bid”), bidders submit their maximum bids to eBay, knowing that “eBay will bid incrementally on your behalf (p. 310) up to your maximum bid, which is kept secret from other eBay users” (statement on eBay's bid submission page). Only outbid maximum bids are published. If we assume for the moment that all bidders submit their bid at the same time, then the “proxy bidding” mechanism implements a second-price sealed-bid auction: the bidder with the highest (maximum) bid wins and pays the second highest (maximum) bid plus one bid increment. However, as the auction is dynamic, bidders might decide to submit their maximum bids incrementally, thereby acting as if in an increasing-price open-bid English auction. Yet we emphasize that the auction mechanism on eBay differs in more than one way from a standard English auction. For one thing, not the last, but the highest bid wins. For another, the auction has a fixed deadline. Still, related to what we predict in English and sealed-bid secondprice auctions, bidding one's own private value on eBay (at some point before the end of the auction) can be— under certain assumptions—part of an equilibrium in undominated strategies (Ockenfels and Roth, 2006). Page 3 of 27 Internet Auctions Whether actual bidding on eBay and other Internet auction markets is in line with theory or not, can be studied in the field or in the laboratory. Laboratory experiments test the predictions of auction theory in a highly controlled environment.12 In such experiments, bidders’ values are often induced (and thereby controlled for) by assigning a specific cash redemption value to each bidder. The bidder who wins the item receives a payoff equal to the redemption value minus the price. Bidders who do not win receive zero. In addition, bidders are typically provided sufficient information about the number of potential bidders and the distribution of the bidders’ values, so that the assumptions made in the theoretical auction models are met. In the following, we discuss some bidding patterns—overbidding, winner's curse, and late bidding—that have been investigated both in the lab and on the Internet. 2.1. Overbidding in Private Value Auctions A robust observation in laboratory experiments is that while bidding in ascending-price private-value auctions usually comes very close to the bidding strategy predicted by theory (bidders stay in the auction until their value is reached, and drop out at that point), bidders exhibit a tendency to bid higher than predicted in other private-value auction formats.13 Several explanations have been put forward in the literature to explain overbidding in privatevalue auctions. The oldest of these explanations is the existence of risk aversion among bidders. A risk-averse bidder will prefer to bid higher in order to increase the likelihood of winning in exchange for a lower profit in case of winning. However, evidence from laboratory experiments suggests that while the assumption of risk aversion can account for some of the overbidding observed in private-value auctions (Cox et al., 1985), it cannot explain the full extent of overbidding and, in particular, not in all auction (p. 311) formats.14 In a more recent study, Kirchkamp et al. (2006), provide a rigorous experimental comparison of first- and second-price auctions, in which the degree of risk is varied. While they find no effect of risk-aversion on bidding in second-price auctions (which corresponds to the fact that the equilibrium in second-price auctions is in weakly dominant strategies that should not be affected by the bidders’ risk attitudes), they identify and quantify a significant effect of risk-aversion on bidding in first-price auctions. Interestingly, the evidence once again suggests that risk-aversion is insufficient to explain the full extent of overbidding. Regret is a complementary explanation of overbidding behavior in laboratory experiments. The idea is that a bidder who loses a first-price auction after submitting a discounted bid, but observes a winning bid lower than one's own value, will regret not having bid higher (“loser-regret”). Similarly, a bidder who wins an auction, but observes a second highest bid (much) lower than the own bid, will regret not having submitted a lower bid (“winner-regret”). Ockenfels and Selten (2005) explore this idea in laboratory sealed-bid first-price auctions with private values. They find that auctions in which feedback on the losing bids is provided yield lower revenues than auctions where this feedback is not given. They introduce the concept of weighted impulse balance equilibrium, which is based on a principle of ex-post rationality and incorporates a concern for social comparison, and show that this captures their results and those of Isaac and Walker (1985) in a related experiment. Filiz-Ozbay and Ozbay (2007) and Engelbrecht-Wiggans and Katok (2007) formalize a similar idea, assuming that bidders anticipate ex ante that they would experience negative utility from post-auction regret and so take the effect into account when submitting their bids. Filiz-Ozbay and Ozbay (2007) conduct laboratory experiments with first-price sealed-bid auctions in which they, too, manipulate the feedback that bidders receive after the auction in order to allow for winner-, loser- or neither form of regret. In particular, in the winner-regret condition, the auction winner is informed about the second-highest bid after the auction, while the losers do not receive any information. In the loser-regret condition, only the losers are informed about the winning bid, while the winners receive no feedback. The results provide strong evidence for loser-regret, but no significant support for winner-regret. With feedback to losers about the winning bid, bidders submit significantly higher bids than with no feedback. A small but significantly negative effect is also found when winners are informed (rather than not informed) about the secondhighest bid. Further evidence comes from an experiment by Engelbrecht-Wiggans and Katok (2009), who allow for learning, and let subjects bid against computerized bidders in order to exclude social determinants of bidding behavior. They implement different payment rules inducing a controlled variation of the variance of payoffs, which allows them to isolate the effect of risk aversion on bidding behavior. While an effect of risk-aversion is not supported by the data, the authors find evidence of both winner-regret and loser-regret. In addition, they find that the magnitude Page 4 of 27 Internet Auctions of regret persists with experience. (p. 312) While, as Filiz-Ozbay and Ozbay (2007) suggest, winner-regret only plays a role in first-price auctions, loser-regret may also be felt in Dutch auctions, where the loser is always informed about the winner's bid, but the winner is typically not informed about the loser's potential bid. So far, however, no attempts have been made in the literature to explain the overbidding that is sometimes observed in second-price or ascending-bid auctions with regret. Another explanation of overbidding that can rationalize overbidding both in first-price and in second-price sealedbid auctions is spite, i.e., the negative utility of seeing someone else win the auction (see Morgan et al., 2003, and Ockenfels and Selten, 2005). Cooper and Fang (2008) provide some experimental evidence that—consistent with the spite hypothesis—bidders in second-price auctions are more likely to overbid if they believe that other bidders have much higher values than themselves. Note, however, that the interpretation of the observations in terms of social comparison may be confounded by the fact that a greater difference between one's own value and that of other bidders also implies a lower risk of being exposed when overbidding. Hence, as the difference between values increases, the expected cost of overbidding decreases for the bidder with the lower value.15 The studies discussed so far concern laboratory experiments testing standard auction theory. There appears to be much less overbidding in Internet auction formats. Garratt et al. (2008) conduct second-price auctions over the Internet inducing values for experienced eBay buyers and sellers. While they find some variance of bids around the induced values, they do not observe a particular tendency to overbid or underbid. Ariely et al. (2005) employ eBay's hybrid format in the laboratory and find that over time—with increased experience of bidders—bids converge to induced values. Greiner and Ockenfels (2012) use the eBay platform itself to conduct an experiment with induced values. This study, too, does not reveal any systematic overbidding or underbidding. On average, the losers’ last bids were very close to their private redemption values (winners’ bids are not observed on eBay). In fact, some studies report significant underbidding due to “reactive” or “incremental” biddings. Zeithammer and Adams (2010), for instance, find a downward bias of bids on eBay due to “some sort of reactive bidding” (see Section 3.3 for more on this). More specifically, their field data suggest that the top proxy bid is often less than the top valuation and too close to the second highest bid. On the other hand, there is evidence on the Internet for a specific kind of overbidding, occurring when there is also competition on the supply side. In a recent study, Lee and Malmendier (2011) collect price data from a large sample of Internet auctions that have a fixed-price offer for an identical item (in exactly the same quality) available on the same website in the course of the entire auction. This fixed-price offer is obviously an easily accessible outside option for bidders and, thus, should define the upper bound for bids in the auction. Surprisingly, however, they find that the auction price is higher than the fixed-price in about 42 percent of the auctions of their first database and in about 48 percent of their second database.16 On average, the winners in the auctions pay about 2 percent more than the fixed-price offer. These observations are in line with, but even stronger than, (p. 313) earlier results by Ariely and Simonson (2003), who report that for 98.8 percent of the 500 online auctions that they study a lower retail price can be found within 10 minutes of search on the Internet. Such phenomena obviously cannot easily be explained by risk-aversion, regret, or social comparison, because the alternative offer comes at no risk and reduces the price to be paid and thus also one's relative position. One alternative explanation might be auction fever.17 The term “auction fever” usually refers to a kind of excitement or arousal of bidders due to the competitive interaction with other bidders. One approach to model the auction fever phenomenon is to assume a quasi-endowment effect (Thaler, 1980), which implies that the value of an item increases during an auction as the bidders become more and more emotionally attached to the item. Auction fever might also be based on a rivalry effect (based on arousal in competition, Heyman et al. 2004), or on uncertain preferences, that is, bidders’ uncertainty in the assessment of their own willingness to pay (Ariely and Simonson, 2003). Some evidence from empirical and experimental studies suggests that both the quasi-endowment and the rivalry effect are present. Ku et al. (2005), for example, aim to distinguish between rational bidding and the different types of auction fever. They collect escalation of commitment, a kind of sunk-cost fallacy (having invested into search and bidding, bidders would like to recover those costs by winning rather than realizing the loss). escalation of commitment, a kind of sunkcost fallacy (having invested into search and bidding, bidders would like to recover Page 5 of 27 Internet Auctions those costs by winning rather than realizing the loss). escalation of commitment, a kind of sunk-cost fallacy (having invested into search and bidding, bidders would like to recover those costs by winning rather than realizing the loss).escalation of commitment, a kind of sunk-cost fallacy (having invested into search and bidding, bidders would like to recover those costs by winning rather than realizing the loss). The authors use data from art auctions, conducted both live and over the Internet, and combine the data with survey responses obtained from (most of) the bidders. They find that the observations are both in line with simple models of item commitment (quasiendowment) and with competitive arousal (rivalry). Furthermore, bidders at live auctions report to have exceeded their bid limits more often than Internet bidders. An additional laboratory experiment provides further support for both explanations. Similarly, Heyman et al. (2004) use a survey on hypothetical Internet auction behavior as well as a performancepaid experiment to distinguish between the rivalry effect (called “opponent effect” here) and the quasi-endowment effect. The former is tested by correlating the final bid to the number of bids others submitted, and the latter by correlating the final bid to the amount of time a participant held the highest bid in the course of the dynamic auction. Again, both explanations find support in the data. Bidders submit higher final bids with more competition and with longer temporary “possession” of the item during the auction. An implication of these results is that lowering start prices might enhance the seller's revenue by increasing the probability or extent of auction fever, because they allow for more rivalry and longer endowment periods. In fact, low start prices are very popular in (p. 314) Internet auctions and the currently highest bidder is often referred to as the “current winner.” Lee and Malmendier (2011) use their field data on overbidding to examine the evidence for two alternative explanations: auction fever versus a limited attention to the outside options. They find almost no correlation between the time spent on the auction and the size of bids and interpret this as evidence against the quasiendowment effect and, thus, as evidence against auction fever. In contrast, they do find some evidence for the limited attention hypothesis, because they observe a strong positive correlation between the on-screen distance of auction listings to the corresponding fixed-price listings and the probability that the auction receives high bids. To conclude, while laboratory research in standard auction formats has produced substantial and robust evidence for overbidding (measured in terms of one's own value), the evidence is mixed in Internet auctions. In simple eBaylike environments, there is so far hardly any evidence for overbidding, and even some evidence for underbidding due to incremental or reactive bidding strategies. When measured with respect to competing offers, however, there is strong evidence that some buyers overpay. The reasons are not fully understood yet. It seems that risk aversion actually accounts for some of the overbidding in laboratory first-price sealed-bid auctions, but it seems to have little relevance in many other auction formats. Ex-post rationality and regret also seems to play a role. However, risk aversion and regret cannot explain paying more than what is offered by other sellers. Auction fever, due to competitive emotional arousal or to a quasi-endowment effect, finds mixed support in experimental and empirical data. Limited attention may also contribute to overbidding, adding yet another behavioral effect to the set of possible explanations. While it seems plausible that some combination of these effects drives overbidding behavior in different auction settings, more research is need to attain a full picture.18 2.2. The Winner's Curse in Common Value Auctions Winning a common value auction is often “bad news,” because it implies that the winner was the most optimistic (had the highest signal) regarding the true value of the good. In fact, abundant experimental (and some empirical) evidence shows that winners often overpay the auctioned good, i.e., fall prey to the winner's curse (Kagel and Levin, 1986, 2002; Thaler, 1988). The greater the number of bidders, the more likely it is for the winner's curse to occur (Bazerman and Samuelson, 1983). While the extent of the phenomenon depends on gender, grades, and other traits (Casari et al. 2007), it is generally persistent over many rounds (Lind and Plott, 1991), and also appears if the good has only a weak common-value component (Rose and Kagel, 2009), proving its general robustness (Grosskopf et al., 2007; Charness and Levi 2009). The evidence for the winner's curse from Internet auctions is weaker than the evidence from laboratory experiments. Bajari und Hortaçsu (2003) analyze a (p. 315) sample of 407 auctions for mint and proof sets of collectible coins on eBay. The authors argue that given the existence of an active resale market for coins and the impossibility to inspect coins before purchase, evaluating these coins involves substantial uncertainty about their common value component. Given the common value character of the items, rational bidders should discount their Page 6 of 27 Internet Auctions bids to prevent the winner's curse, and the more so the higher the number of bidders participating in the auction. Using both Poisson regressions and a structural model incorporating bidding behavior in second-price commonvalue auctions, Bajari und Hortaçsu (2003) find a significant negative impact of the number of participating bidders in the coin auctions on the size of individual bids. In particular, the structural model yields a significant degree of uncertainty concerning the value of the good and shows that with each additional bidder bids are reduced by 3.2 percent on average.19 The model also indicates that a reduction of the uncertainty concerning the item's value would lead to a significant increase in the bids relative to signal—as theory predicts. These results suggest that bidders in the field are aware of the winner's curse and rationally adapt their behavior in order to prevent it. Bajari and Hortaçsu (2003) do not find that inexperienced bidders are more susceptible to the winner's curse, but a study by Jin and Kato (2006) does. They report a strong positive correlation between bidder's inexperience and the degree to which the item is overpaid. They first gathered data on bids and prices for ungraded baseball cards that were traded on eBay and found that buyers pay 30 percent to 50 percent more when the seller claims the card to be of high quality. In a second step, the authors purchased 100 ungraded baseball cards from eBay, half from seller's who made claims about high quality, and half from sellers who did not make such claims. The researchers carried out their bidding in a way as to minimize distortions of the data: they only bid in the last five minutes and if there was at least one bid present. After purchase, all cards were submitted to a professional grading service, in order to obtain an objective measure of item quality. It turns out that—conditional on authentic delivery—the quality of cards received from high-claim sellers is indistinguishable from the quality of cards purchased from the other sellers. When including defaults and counterfeits as zero-quality deliveries, the average quality from sellers claiming a high quality is actually lower than from sellers who made no claims. Since especially inexperienced bidders pay significantly higher prices for the high-claim cards, it seems that they fall prey to the winner's curse by being too optimistic about the reliability of the claims. The most optimistic bidder in each auction wins the item and more often than not pays a too high price.20 A novel approach to assess the role of information dispersion in Internet auctions is introduced by Yin (2009), who complements the information available from the online auction platform with survey information on the distribution of value assessments. The survey participants (not involved in the auctions) are asked to assess the maximum value of the items after receiving the complete online item descriptions, but no information on the seller or other aspects of the auction. Yin (2009) uses the resulting distribution of values to estimate rational (p. 316) equilibrium and naive bidding strategies both under the assumption of common and private values. She concludes that her data on PC auctions are best in line with the equilibrium of a common value setting with almost no signs of a winner's curse. The result hinges on the fact that bids negatively correlate with the dispersion of the value assessments, because bidders bid more cautiously, avoiding the winner's curse and inducing lower prices the greater the dispersion is. This result, too, indicates that bidder behavior in Internet auctions may be more sophisticated than laboratory studies have suggested. More research seems necessary to identify what behavioral, institutional and methodological differences drive the different conclusions regarding bidder naivety and winner's curse in the laboratory and in Internet auctions. 2.3. Late and Incremental Bidding In most traditional auction formats, the timing of bids is not an issue. In sealed-bid auctions, all bidders submit their bid simultaneously. In increasing or decreasing price clock auctions, the auction is paced by the auctioneer and the size of the bids. However, in English auctions and its dynamic variants on Internet platforms, the bidders endogenously drive the auction price, turning the timing of the bids into part of the bidding strategy. In this setting, late bidding is a pervasive feature of Internet auctions (e.g., Roth and Ockenfels, 2002; Bajari and Hortaçsu, 2003). The theoretical and empirical work on this phenomenon suggests that it is closely related to the ending rule of the respective auction. The simplest rule for ending a dynamic auction is a hard close, that is, when the prespecified end of the auction duration is reached, the auction terminates and the bidder with the highest bid is the winner of the auction. This rule is employed on eBay, among other sites. Based on survey data, Lucking-Reiley (2000a) lists Internet auction lengths ranging from 60 minutes up to 90 days. Other platforms, such as Amazon, Yahoo and uBid auction platforms use or used (not all auction houses still exist) a flexible timing rule often referred to as soft close: whenever a bid is submitted in the last, say, 10 minutes of the soft-close auction, the auction time is extended by a Page 7 of 27 Internet Auctions fixed amount of time, for example, another 10 minutes. While a substantial amount of late bidding (also called “sniping”) is found on hardclose auctions, very little late bidding is observed on soft-close auctions. Figure 12.1 from Roth and Ockenfels (2002) reports the cumulative distribution of the timing of the last bid in a sample of 480 eBay and Amazon auctions with at least two active bidders. About 50 percent of the eBay hard-close auctions attracted bids in the last 5 minutes (37 percent in the last minute and 12 percent in the last 10 seconds) compared to only about 3 percent of the Amazon soft-close auctions receiving bids in the last 5 minutes before the initially scheduled closing time or later.21 Figure 12.1 Cumulative Distributions of Timing of Auctions’ Last Bids Timing on eBay and Amazon (from Roth and Ockenfels, 2002). In a controlled laboratory experiment with a private-value setting, Ariely et al. (2005) replicate the empirical findings. While bidders in the Amazon-like mechanism converge to bidding early, participants in the eBaymechanism tend to bid later. (p. 317) There are many good reasons for bidding late. The difference in bid timing between eBay's hard-close and Amazon's soft-close design suggests that sniping is driven by strategic behavior rather than simple non-strategic timing (Roth and Ockenfels, 2002).22 In particular, sniping avoids at least four different types of bidding wars: with like-minded bidders, with uninformed or less informed bidders, with shill-bidding sellers, and with incremental bidders. Ockenfels and Roth (2006) develop a game-theoretical model with which they demonstrate that, given a positive probability that a bid does not arrive when submitted in the last minute, mutual late bidding might constitute an equilibrium between bidders who have similar values for the good (i.e., “like-minded” bidders). The reason is that when submitting early, the bidder either loses the auction (yielding a profit of zero) or wins the auction (yielding a profit close to zero, since the other like-minded bidders submit similarly high bids). But if all bidders submit late, then there is a positive probability that only one of the bids arrives, implying a high profit for the winner. However, this explanation of sniping is not fully supported by the empirical evidence. For instance, Ariely et al. (2005) find that, contrary to the theory, changing the arrival probability of late bids from 80 percent to 100 percent leads to an increase in the extent of late-bidding in their experiment.23 Bidding late might also be a reasonable strategy if values are interdependent (i.e., have a common value component). Imagine an expert spots a particularly good deal on eBay, an original antique which is wrongly described by the seller. If the expert would go into the auction early, he runs the risk of revealing this information, as other bidders may observe his activity. Thus, he has strong incentives to wait until the end (see Bajari and Hortaçsu, 2003, and Ockenfels and Roth, 2006, for formalizations of this idea). Sniping may also be an effective strategy against shill-bidding. If the seller bids himself in the auction in order to raise the price close to the maximum bid of the highest bidder, then delaying the bid until the end of the auction undermines that (p. 318) strategy effectively, as it does not leave the seller enough time to approach the real maximum bid (Barbaro and Bracht, 2006; Engelberg and Williams, 2009). Late bidding might be a best response to a different type of bidding commonly observed in Internet auctions: incremental bidding. Here, a bidder does not make use of the proxy mechanism provided by the auction platform, but rather increases his “maximum” bid incrementally—as if bidding in an English auction. Against incremental bidders, waiting and sniping prevents a bidding war (e.g., Ely and Hossain, 2009). The empirical evidence for the existence of incremental bidding and sniping as a counter-strategy is quite strong. Among others, Ockenfels and Roth (2006) report the co-existence of incremental and late bidding on eBay. Wintr (2008) observes substantially later bids from other bidders in the presence of incremental bidders in the auction. Similarly, Ely and Hossain (2009) find their model of incremental and late bidding confirmed in the data collected in a field experiment. See Page 8 of 27 Internet Auctions also Ariely et al. (2005), Ockenfels and Roth (2006), Wilcox (2000), Borle et al. (2006) and Zeithammer and Adams (2010), who all find evidence that incremental bidders tend to be more inexperienced. When some bidders use late bidding in response to incremental bidding, why is there incremental bidding in the first place? A number of attempts to explain the incremental bidding pattern in Internet auctions have been made in the literature. Proposed explanations include auction fever or quasi-endowments (see above), as well as bidding simultaneously in multiple competing auctions. The auction fever explanation is straightforward. Bidders are assumed to become emotionally so involved in the auction that the perceived in-auction value of the item (or of winning the auction) increases in the course of the auction (see Section 3.1). Bidders facing competing auctions are assumed to bid incrementally to avoid being “caught” with a high price on one auction, while the competing auctions end with lower prices. In fact, several authors have shown that bidding incrementally across auctions (always placing the next bid at the auction with the currently lowest price) can be an equilibrium in a competing auction setting (Anwar et al., 2006; Peters and Severinov, 2006).24 Similarly, in Nekipelov's (2007) model of multiple concurrent auctions, bidders have an incentive to bid early to deter entry, which might also lead to incremental bidding patterns. Finally, incremental bidding may also be due to some form of uncertain preferences. Rasmusen (2007), for example, proposes a model in which bidders are initially not fully aware of their own value. But, the bidders only invest into thinking hard and gathering information as the auction proceeds and they recognize that they might actually have a chance to win. To avoid negative surprises, bidders start with low bids and increase these only as good news on the true value arrives. Similarly, Hossain (2008) develops a model of a bidder who only knows whether his value exceeds a given price once he sees the price. Hossain shows that such a bidder would start with a low bid and then bids incrementally. In a recent theoretical paper Ambrus and Burns (2010) show that, when bidders are not able to continuously participate in an auction but can only place bids at random times, there exist equilibria involving both incremental and late bidding. Regarding the profitability of late bidding, the field evidence is mixed. (p. 319) Bajari and Hortaçsu (2003) and Wintr (2008), for example, find no differences in final prices paid by early or late bidders, while Ely and Hossain (2009), Gray and Reiley (2007) and Houser and Wooders (2005) find small positive effects of sniping on bidder profits. To sum up, there is strong evidence both for the existence of incremental and of late bidding in Internet auctions. The phenomena seem to be driven partly by the interaction of naive and sophisticated bidding strategies. Also, the design of the auction termination rules proves to be an important strategic choice variable of Internet auction platforms, strongly affecting the timing of bids. However, the evidence on the extent to which incremental and late bidding affects efficiency and revenue is mixed, although there are indications that sniping tends to hamper auction efficiency and revenue. Ockenfels and Roth (2010) provide a more detailed survey of the literature on sniping and ending rules in Internet auctions. 3. Seller Strategies 3.1. Auction Formats Internet auction platforms can choose which auction formats to offer, and sellers and buyers can choose which platform to use. With so many choices, the question obviously is which designs are successful in which environments. The Revenue Equivalence Theorem that we discussed in Section 2 suggests that auction design may not matter much, given a few simple assumptions are met. There is quite a bit of empirical and experimental evidence, however, that even in very controlled environments revenue equivalence does not hold.25 In the first empirical Internet auction study, Lucking-Reiley (1999) compares the revenue of different mechanisms for online auctions in a field experiment conducted on an Internet newsgroup. Auctions on that platform were common so that different mechanisms could be easily tested. The author auctioned identical sets of role-playing game cards to experienced bidders, who were familiar with the items. The time horizon of the auctions was much longer than in related laboratory studies (days rather than minutes). In the first experiment, Lucking-Reiley (1999) sold 351 cards using first-price and Dutch auctions, controlling for order effects. The main result is that for 70 percent of the card sets, the Dutch mechanism yielded higher revenues than the first-price auction, with an average price advantage of 30 percent. This result is both in contrast to auction theory, which predicts identical revenues across these formats for any risk preferences, and to the findings from laboratory experiments, in which usually higher revenues are observed in first-price auctions. Page 9 of 27 Internet Auctions In the second experiment, Lucking-Reiley (1999) sold 368 game cards in English and second-price auctions. This time he observes an order effect. The auction type conducted first (be it English or second-price auction) yields lower revenues than (p. 320) the second auction type. However, the overall differences in revenues are small and barely significant. Given the small differences in observed revenues, Lucking-Reiley (1999) conjectures that English auctions and second-price auctions are approximately revenue equivalent in the field. The design of Internet auctions, however, eludes most traditional formats, patching auction types together or designing completely new formats. Formats such as the unique bid auctions, which we discuss in detail in a later subsection, or the combination of proxy-bidding and buy-it-now pricing took over the market long before any auction theorist had a chance to study them. In fact, excellent auction theory reviews (e.g., Klemperer 1999 or Milgrom 2004) are outpaced so quickly by the inventiveness of Internet auction designers that we must concede that auction research is transforming into history faster than we can write it up. On the other hand, however, auction researchers have also come up with formats that are still awaiting an application—or at least a serious test—in the field. Katok and Kwasnika (2008), for example, present an intriguing study on the effects that can be achieved by changing the pace of auctions. Comparing first-price sealed-bid auctions with Dutch auctions, they find that the former yield greater revenues when time pressure is high. Interestingly, the opposite is true, when the clock runs slowly. In another study that proposes that auction duration may be an important design variable, Füllbrunn and Sadrieh (2012) compare standard hard-close ascending-bid open auction with candle auctions that only differ in the termination mechanism. They find that using a stochastic termination rule as in candle auctions can substantially speed up the auction without a loss of efficiency, revenue, or suspense. These examples show that the research on auction formats is still far from being concluded, especially because new design parameters are being introduced frequently. 3.2. Reserve Prices Auction theory postulates that expected revenues in an auction can be maximized by setting an appropriate reserve price or minimum bid (Section 2). Reiley (2006) uses a field experiment to analyze the effects of reserve prices in Internet auctions. He focuses on three theoretical predictions: First, increasing the reserve price should decrease the number of bidders participating in the auction. Second, a higher reserve price should decrease the probability of selling an item. And finally, increasing the reserve price should increase the revenue, in case the auction is successful. The field experiment employed sealed-bid, first-price auctions with an Internet news group for game cards, rather than a dedicated auction platform such as eBay. The reserve price in the auctions was used as the main treatment variation. As a reference for induced reserve prices, Reiley (2006) used the so-called Cloister price, which equals the average selling price for the different game cards on a weekly basis. (p. 321) Reiley (2006) finds that, as theory predicts, introducing a reserve price equal to 90 percent of the reference price decreases the number of bidders and the number of bids. Varying the reserve price yielded a monotonic linear relation between reserve price and number of bidders. In particular, increasing the reserve price from 10 percent to 70 percent of the reference price decreased the number of bids from 9.8 to 1.0 on average (with no further significant change for higher reserve prices). Correspondingly, the probability of selling an item is reduced with higher reserve prices. Finally, auction revenues for the same card, conditional on auction success, are significantly higher if a reserve price of 90 percent was set, compared to no reserve price.26 However, considering unconditional expected auction revenues (i.e., assuming that an unsold card yielded a price of $0) draws an ambiguous picture: in pair-wise comparison, auctions with a reserve price of 90 percent do not yield higher revenues than those without a reserve price, as many cards were not sold with the reserve price present. The net profits were lowest with too low reserve prices of 10 percent to 50 percent (with expected revenue of about 80 percent of Cloister value) and too high reserve prices of 120 percent to 150 percent (yielding about 90 percent of Cloister value, on average). Intermediate reserve prices between 80 percent and 100 percent resulted in the highest expected revenues, about 100 percent of Cloister value on average. Other field experiments that test the variation of the public reserve price in richer Internet auction settings find similar results. Ariely and Simonson (2003) compare high and low minimum prices in high supply (“thick”) and low supply (“thin”) markets. They report that the positive effect of the reserve price is mediated by competition, that is, reserve prices are hardly affective when the market is competitive, but do well in increasing revenue when the market is thin. While the result is well in line with the theoretical prediction in the reserve price competition model Page 10 of 27 Internet Auctions by McAfee (1993), the authors conjecture that bidders are more likely to “anchor” their value estimates on the starting price, when the market is thin and only few comparisons are possible than in highly competitive markets. Obviously, the anchoring hypothesis requires some form of value uncertainty or bounded rationality on the bidders’ part. The phenomenon may be connected to the framing effect observed by Hossain and Morgan (2006). Selling Xbox games and music CDs on eBay, they systematically vary the reserve price and the shipping and handling fees. For the high-priced Xboxes, they observe that total selling prices (i.e., including shipping fees) are highest when the reserve price is low and the shipping fee is high. Since they do not observe the same effect for the low-priced music CDs, they conjecture that the (psychological) effect of reserve prices on bidding behavior is rather complex and depends on a number of other parameters of the item and the auction environment. Some Internet auction platforms allow sellers to set their reserve prices secretly rather than openly. In particular for high-value items such secret reserve prices seem to be very popular (Bajari and Hortaçsu, 2003). Theoretically, in a private value setting with risk neutral bidders, hiding a reserve price does not have positive effects on the seller's auction revenues (Elyakime et al., 1994; Nagareda, 2003). (p. 322) However, secret rather than open reserve prices may have positive effects if bidders are risk-averse (Li and Tan, 2000), have reference-based preferences (Rosenkranz and Schmitz, 2007), or are bidding in an affiliated-value auction (Vincent, 1995). Estimating a structural econometric model with eBay data on coin auctions, Bajari and Hortaçsu (2003) find that an optimal secret reserve price increases expected revenues by about 1 percent. Katkar and Reiley (2006) find a contradicting result. They implement a field experiment to compare the effect of secret versus public reserve prices. They auction 50 pairs of identical Pokemon trading cards on eBay. One card of each pair is auctioned with a public reserve price at 30 percent of its book value and the other with the required minimum public reserve price and a secret reserve at 30 percent of the book value. The auctions with secret reserve prices attract fewer bidders, are less likely to end successfully, and result in an about 9 percent lower final auction price in the case of success. The negative effect of secret reserve prices is confirmed by an empirical study on comic books sales on eBay. Dewally and Ederington (2004) find that auctions using a secret reserve price (that is known to exist, but is secret in size) attract fewer bidders and generate less revenue for the seller. 3.3. Shill Bidding A shill bid is a bid placed by the seller (or a confederate) without being identified as a seller's bid. A shill bid can be differentiated from other “normal” bids by its purpose, which obviously is not for the seller to buy the own property (especially not, since buying the own property incurs substantial trading fees on most auction platforms). Shill bidding, in general, has one of two purposes. On the one hand, a shill bid may be used much as a secret reserve price, allowing the seller to keep the item, if the bids are not above the seller valuation. On the other hand, shill bids may be used dynamically to “drive up” (or to “push up”) the final price, by incrementally testing the limit of buyers’ willingness to pay. While both types of shill bidding are forbidden on most platforms (and legally prohibited by some jurisdictions), the reasons for prohibiting them are very different. The reserve price shill bids are prohibited because using them evades the high fees that auction platforms ask for installing proper reserve prices. The dynamic shill bids are prohibited to protect the buyers’ rent from being exploited by sellers.27 Fully rational bidders, who are informed that sellers can shill bids, will take shill bidding into account and adjust their bidding behavior. Depending on the setting, this may or may not result in loss of efficiency and a biased allocation of rent. Izmalkov (2004), for example, shows that in an independent private value model with an ascending open auction, the shill bidding equilibrium is similar to Myerson's (1981) optimal auction outcome, which relies on a public reserve price. For the common-value setting, Chakraborty and Kosmopoulou (2004) demonstrate (p. 323) that bidders, anticipating shill bidding, drop their bids so low that the sellers would prefer to voluntarily commit to a noshilling strategy. The intuition underlying both results is simply that strategically acting bidders will adjust their bids to take the presence of shill bidding into account. In an experimental study on common-value second-price auctions, Kosmopoulou and de Silva (2007) test whether buyers react as theoretically predicted when sellers are allowed to place bids, too. Buyers are either informed or not about the bidding possibility for the seller. The results show that while there is general overparticipation in the auction and the number of bidders is not affected by the opportunity to shill-bid, the bidders, who are aware of the Page 11 of 27 Internet Auctions seller's shill-bidding opportunity, bid more carefully than the others. This leads to lower prices and positive bidder payoffs even when sellers use shill bidding. Identifying shill bidding in the field is difficult, because the seller can register with multiple identities or employ confederates to hide their shill bids.28 One way of identifying shill bids is to define criteria that are more likely for shill bids than for normal bids. Using this approach on a large field data set, Kauffmann and Wood (2005) estimate that (if defined very narrowly) about 6 percent of eBay auctions contain shill bids. They also find that using a shill bid positively correlates with auction revenues. A positive effect on revenue, however, is not reported by Hoppe and Sadrieh (2007), who run controlled field experiments comparing public reserve prices and reserve price bids by confederates. Estimating the optimal reserve price from the distribution of bids, they show that the prices in successful auctions are higher both with estimated optimal public reserve prices and shill bids as compared to the auctions with the minimum reserve price. The only positive effect of bid shilling that they can identify for the sellers concerns the possibility to avoid paying fees for public reserve prices. 3.4. Buy-it-Now Offers A buy-it-now offer is the option provided by the seller to buyers to end the auction immediately at a fixed price. Such a feature was introduced on yahoo.com in 1999, and eBay and other auction sites followed in 2000 (LuckingReiley, 2000a). Buy-it-now offers are popular among sellers: about 40 percent of eBay sellers use the option (Reynolds and Wooders, 2009). There are two different versions of the buy-it-now feature used on Internet platforms: either the fixed-price offer disappears as soon as the first auction bid arrives (e.g., on eBay), or it remains valid until the end of the auction (e.g., on uBid, Bid or Buy, or the Yahoo and Amazon auctions). At first glance, the popularity of buy-it-now offers is puzzling: the advantage of an auction lies in its ability to elicit competitive prices from an unknown demand. Posting a buy-it-now offer (especially a permanent one) puts an upper bound on the price which might be achieved in an auction. One explanation for the existence of buy-it now offers is impatience of the economic agents. Obviously, if bidders (p. 324) incur costs of waiting or participation, then they might prefer to end the auction earlier at a higher price. If sellers are impatient, too, then this will even reinforce the incentives to set an acceptable buy-it-now offer (Dennehy, 2000). A second explanation is risk-aversion on the side of the buyer or the seller. If bidders are risk averse, they would prefer to accept a safe buy-it-now offer rather than to risk losing the auction (Budish and Takeyama, 2001; Reynolds and Wooders, 2009). This can hamper auction efficiency if bidders have heterogeneous risk preferences (as more risk-averse bidders with lower values might accept earlier than less risk-averse bidders with higher values), but not if they have homogenous risk-preferences (Hidvégi, Wang and Whinston, 2006). Mathews and Katzman (2006) show that while a risk-neutral seller facing risk-neutral bidders would not use a temporary buy-itnow option, a risk-averse seller may prefer selling using the buy-it-now option. The authors furthermore show that under certain assumption about the bidders’ value distribution, the buy-it-now option may actually be paretoimproving, that is, make both the bidders and the seller better off by lowering the volatility of the item's price. Empirical studies (e.g., Durham et al., 2004, Song and Baker, 2007, and Hendricks et al., 2008) generally find small positive effects of the buy-it-now option on the total revenue in eBay auctions. In particular, experienced sellers are more likely to offer a buy-it-now price and buy-it-now offers from sellers with good reputations are more likely to be accepted. There, however, are also some conflicting results. For example, Popkowski Leszczyc et al. (2009) report that the positive effect of buy-it-now prices is moderated by the uncertainty of the market concerning the distribution of values. In laboratory experiments, risk-aversion may be a reason for offering or accepting the buy-it-now option, but impatience can be ruled out, since participants have to stay for a fixed amount of time. Seifert (2006) reports less bidding, higher seller revenues, and lower price variance when a temporary buy-it-now option is present and there are more than three bidders in a private value auction. Shahriar and Wooders (2011) study the effect of a temporary buy-it-now option on seller revenues in both a private and a common value environment. In the former, revenues are significantly higher and have a lower variance with a buy-it-now option than without, as predicted theoretically in the presence of risk-averse bidders. For the common value setting, standard theory neither predicts an effect of the buy-it-now option in the case of risk-neutral, nor in the case of risk-averse bidders. The authors provide a “winner's curse” account that can explain the observed (small and insignificant) revenue increase with Page 12 of 27 Internet Auctions the buy-it-now option. Peeters et al. (2007) do not find a strong positive effect of the temporary buy-it-now option in their private value auctions. But, they discover that rejected buy-it-now prices have a significant anchoring effect on the prices in the bidding phase, with bids rarely exceeding the anchor. Grebe, Ivanova-Stenzel, and Kröger (2010) conduct a laboratory experiment with eBay members, in which sellers can set temporary buy-it-now prices. They observe that the more experienced the bidders are, the higher the buy-it-now price is that the sellers choose. This they attribute to the (p. 325) fact that more experienced bidders generally bid closer to their true values, thus increasing the auction revenue compared to less experienced bidders. Hence, the more experienced the bidders, the higher the seller's continuation payoff (i.e., the payoff from the bidding stage) and, thus, the higher the chosen buy-it-now price. All in all the literature on the buy-it-now price is growing steadily, but the results remain at best patchy. It seems clear that impatience and risk-aversion on both market sides are driving forces of the widening usage of buy-it-now prices. However, there is also some limited evidence that rejected buy-it-now prices may be used as anchors for the following bidding stage, certainly limiting the bids, but perhaps also driving them towards the limit. 3.5. Bidding Fees and All-Pay-Auctions An interesting recent development in Internet auctions is the emergence of platforms which charge bidding fees. These auctions share features with all-pay auctions known in the economics literature. Generally, in all-pay auctions, both winning and losing bidders pay their bids, but only the highest bidder receives the item (see Riley, 1980, Fudenberg and Tirole, 1986, and Amann and Leininger, 1995, 1996, among others, for theoretical contributions). Similarly, in the auctions conducted on Internet sites like swoopo.com, a bidding fee is to be paid for each bid, and independent of the success of that particular bid. In particular, swoopo.com auctions start off at a price of $0. Like on eBay, bidders can join the auction at any time and submit a bid. After each bid, the ending time of the auction is reset, implementing a soft-close mechanism similar to the one for Amazon auctions discussed in our section on sniping behavior above. The extension time decreases with the current price, with a final time window of 15 seconds only. There is a prescribed nominal bid increment of typically a penny ($0.01, which is why these auctions are often called “penny auctions”). However, there is also a fee connected with the submission of each bid, usually equaling $0.75.29 When a penny auction ends as the time window elapses without a new bid, the bidder who submitted the last bid increment wins the item and pays the final price p. But all participating bidders have paid their bidding fees, so the seller (platform) revenues are equal to the final item price p plus all bidding fees (equaling 75p with bid increments of $0.01 and bidding fees of $0.75), which means that revenues can be quite substantial. On the other hand, the final price p is usually much lower than the average resale price of the item, and auction winners may indeed make substantial surpluses, in particular if they joined the auction late. In the following we focus on two recent papers by Augenblick (2010) and Ockenfels et al. (2010), which combine theoretical and empirical work on all-pay auction both based on data from the platform swoopo.com.30 Both papers set up theoretical models to analyze the most important features of the specific auction mechanism used by swoopo.com. These models assume n bidders for whom the (p. 326) item at auction has a common value. The game is modeled for discrete periods of time and with complete information, with each point in time t representing a subgame. Ockenfels et al. (2010) compute the pure-strategy subgame-perfect equilibria of the swoopo game with homogenous bidders and find that not more than one bid is placed. However, there also exist mixed-strategy equilibria in which all bidders bid with positive probabilities that decrease over time. Augenblick (2010) focuses his analysis on the discrete hazard rates in the auction game, that is, the expected probabilities that the game ends in a given period. He first shows that no bidder will participate once the net value of the item is lower in the next period than the cost of placing a bid. Then, Augenblick (2010) identifies an equilibrium in which at least some players strictly use mixed strategies. In both models, bidders are more likely to engage in bidding if they are subject to a naïve sunk cost fallacy that increases their psychological cost of dropping out the longer they participate in the auction. Ockenfels et al. (2010) analyze data of 23,809 swoopo auctions. Augenblick (2010) has collected data on 166,000 swoopo auctions and 13.3 million individual bids. The most obvious and important insight from the field data in both studies is extensive bidding. On average, each auction attracts 534 and 738 bids, respectively, in the two datasets. On the one hand, final prices are on average only 16.2 percent of the retail reference price (Ockenfels et Page 13 of 27 Internet Auctions al., 2010). This underlines that auction winners can strike a bargain, provided they do not submit too many bids themselves. On the other hand, the seller makes substantial profits, with total auction revenues on average between 25 percent (Ockenfels et al., 2010) and 50 percent (Augenblick, 2010) higher than retail prices. In a recently emerged new variant of the all-pay auction, the winner is not determined by the highest bid, but by the lowest unique bid submitted in the auction.31 The auction runs for a predetermined period. Each bid in the auction costs a certain fee, but in contrast to the penny auctions the bid in the unique lowest bid auction is not a fix increment on the current bid. Instead, the winning bid is the lowest amongst the unique bids. Bidders may submit as many bids as they wish. Gallice (2009) models the unique lowest bid auction as an extensive form game with two or more bidders, a seller, and a fixed duration in which the buyers can submit their costly bids. Gallice (2009) points out that the game does not yield positive profits for the seller when bidders are rational. But when bidders are myopic, the unique bid auction is profitable, because—as in all other all-pay auctions—the auction winner and the seller can both earn substantially by splitting the contributions of the losing bidders. Eichberger and Vinogradov (2008) also present a model of unique lowest bid auctions that allows for multiple bids from each individual. Unlike the model presented by Gallice (2009), they model the unique bid auction as a oneshot game, in which each bidder submits a strategy specifying a set of bids. Pure strategy equilibria only exist in this game if the cost of bidding is very high relative to the item's value. These equilibria are asymmetric, with only one player submitting a positive bid. For the more realistic case of moderate or low cost of bidding, Eichberger and Vinogradov (2008) derive the mixed strategy equilibria of the game with a restricted strategy space. Allowing bidders only to choose strategies that specify bids on “all values up to x,” they show (p. 327) that in the unique symmetric mixed strategy equilibrium of the restricted game, the probability of choosing a strategy decreases both in the number of bidders and in x for all values of x greater than zero.32 Comparing their theoretical results to the data from several unique bid auctions in the United Kingdom and Germany, Eichberger and Vinogradov (2008) find that the downward slope of the distribution of observed bids is in line with their equilibrium predictions, but the observed distributions can only be accounted for by assuming bidder heterogeneity.33 In two related studies, Rapoport et al. (2009) and Raviv and Virag (2009) study one-shot unique bid auctions in which the bidders can only place a single bid each. Rapoport et al. (2009) study both lowest and highest unique bid auctions, while Raviv and Virag (2009) study a highest unique bid auction that differs from the others in that only the winner pays the own bid, while all losers pay a fixed entry fee. Despite the many subtle differences across models and across theoretical approaches, the equilibria in these one-shot single-bid games are very similar to those of the multiple-bid auctions discussed above. As before, the symmetric equilibria are in mixed strategies with probabilities that monotonically decrease in bids for the lowest—and increase for the highest—unique bid auction. The empirical findings of Raviv and Virag (2009) are well in line with the general structure of the predicted equilibrium probabilities. Unique bid and other all-pay auctions seem to be a quickly growing phenomenon in Internet auctions. It appears, however, that these institutions are not primarily designed to allocate scarce resources efficiently (theory suggests they do not). Rather, the suspense of gambling and other recreational aspects seem to dominate the auction design considerations. Such motivational aspects of participation are not yet well understood and require more research. 4. Multi-Unit Auctions From the perspective of auction theory, as long as each bidder only demands a single unit, the theoretical results obtained for the single-unit auctions are easily generalized to the multi-unit case. Multi-unit demand, however, typically increases the strategic and computational complexities and often leads to market power problems. 4.1. Pricing Rules with Single-Unit Demand For multi-unit auctions we can distinguish four standard mechanisms analogous to the single-unit case. Items can be sold in an open auction with an increasing price that stops as soon as the number of bids left in the auction matches the number of items offered (i.e., as soon as demand equals supply). Items can be sold in a decreasing Page 14 of 27 Internet Auctions price auction, where a price clock ticks down and bidders may buy units (p. 328) at the respective current price as long as there are units on supply. A “pay-as-bid” sealed-bid auction can be employed, in which each successful bidder pays the own bid for the unit demanded. Finally, items can be offered on a uniform-price sealedbid auction, where all successful bidders pay the same price per unit. As long as each bidder only demands one unit, multi-unit auction theory typically predicts (under some standard assumptions) that the auction price equals the (expected) n+1th-highest value across bidders, with the n bidders with the highest values winning the auction. The reason is that in an auction where each winner has to pay his bid (“pay-as-bid” sealed-bid or decreasing clock), winners would regret having bid more than necessary to win, that is, having bid more than one increment above the highest losing bid. Thus, game-theory predicts that bidders submit their bids such that the price they will have to pay does not exceed the expected n+1th-highest bid. Since this is true no matter in which auction mechanism they are bidding, the seller on average will obtain the same revenues in all formats.34 The multi-unit price rule most commonly implemented on Internet platforms like eBay states that the highest n bidders win and pay a price equal to the nth highest bid. In other words, the lowest winning bid determines the price. Kittsteiner and Ockenfels (2008) argue that, because each bidder may turn out be the lowest winning bidder, bidders strategically discount their bids in this auction, inducing risks and potential efficiency losses. A more straightforward generalization of eBay's hybrid single-unit auction format is to set the price equal to the n+1thhighest bid, that is, equal to the highest losing bid. This pricing rule has the advantage that it reduces the variance in revenues, increases the number of efficient auction allocations, and makes bidding the own value a (weakly) dominant strategy. Cramton et al. (2009) experimentally study both pricing rules and find that the nth-bid pricing rule yields higher revenues than the n+1th-bid rule. However, in a dynamic auction like eBay, the last bidder who places a bid will be able to influence the price, and therefore strategize. Based on such arguments, eBay changed its multi-unit design from an nth- to an n+1th-highest-bid pricing rule, before it dropped the multi-unit auction platform altogether (Kittsteiner and Ockenfels, 2008).35 4.2. Ad Auctions Among the largest multi-unit auctions in the Internet are the ones used by search engines to sell keyword-related advertisements in search results, such as Google's “AdWords Select” program. Edelman et al. (2007) and Varian (2007, 2009) investigate this specific type of position auction. While in the early days of the Internet advertisement space was sold in bilateral markets, the industry was revolutionized when GoTo.com (later renamed Overture) started using large-scale multi-unit auctions to sell search word ads in 1997. In 2002 Google followed suit, but replaced Overture's simple highest-bid-wins format by an auction that combined the per-click price and the expected traffic (originally measured by the “PageRank” (p. 329) method that was patented by Google) into an aggregate average-per-period-income bid.36 This rather complex mechanism has become even more complicated over time, with one key element of the current system being the “Quality Score,” which combines a number of criteria including historic click-through-rates and the “page quality” (that basically is an enhanced PageRank). Google keeps the exact calculations of the measures secret in order to reduce strategic adaptation and fraud (Varian, 2008). The “Generalized Second Price” (GSP) auction is the original position auction format that was first used by Overture. Basically, whenever an Internet user submits a search query, the search engine parses the search phrase for keywords and conducts a quick auction for advertisement positions on the results page, using presubmitted bids from advertisers. The bidder with the highest bid (per user click) receives the first (most popular) position, the second-highest bidder the next position, and so forth. Under the rules of the Generalized Second Price auction each bidder only pays the bid of the next-highest bidder (plus one increment). Thus, the highest bidder only pays the second-highest bid, the second-highest bidder pays the third-highest bid, etc., until the bidder receiving the last advertisement position slot pays the highest unfulfilled bid.37 With only one advertisement slot this pricing rule corresponds to the regular single-unit second-price auction. At first glance, the popularity of the GSP may seem surprising given that another mechanism, the Vickrey-ClarkeGroves mechanism (VCG), exists that has a unique equilibrium in (weakly) dominant strategies. In the VCG mechanism, each bidder pays the externality imposed on the other bidders by occupying a specific advertisement slot. That is, each bidder pays the difference between the total value all other bidders would have received, if the Page 15 of 27 Internet Auctions bidder would have not participated, and the total value all other bidders receive in the current allocation. In ad position auctions, a bidder's participation shifts all subsequent bidders one slot lower in the list. Thus, in the VCG, each bidder pays the sum of value differences between the bid for the current slot and the bid for the next slot up. In contrast, the GSP auction has multiple equilibria and no unique equilibrium in dominant strategies. To derive a useful prediction for the GSP auction, Edelman et al. (2007) and Varian (2007) introduce the notion of “locally envy-free” Nash equilibria and seek the rest-point in which bidder behavior stabilizes in infinitely repeated GSP auctions. In such an equilibrium, no bidder has an incentive to swap bids with the bidder one position above or below. Both theoretical papers demonstrate the existence of a “locally envy-free Nash equilibrium” in the GSP auction, in which the position allocation, the bidder's payments, and the auction revenues correspond to the dominant equilibrium of the VCG mechanism. This equilibrium is the most disadvantageous amongst all locally envy-free equilibria for the seller, but it is the best for the buyers. Edelman et al. (2007) furthermore show that this equilibrium corresponds to the unique perfect Bayesian equilibrium when the GSP auction is modeled as an English auction instead. In sum, multiple equilibria that do not impose truth-telling exist in GSP auctions. Some of these equilibria, however, are locally envy-free. One of these is particularly (p. 330) attractive to bidders, as their payoffs in the equilibrium coincide with those in the dominant strategy-equilibrium of the VCG mechanism and the unique perfect equilibrium of a corresponding English auction. Moreover, all other locally envy-free equilibria of the GSP lead to lower profits for bidders and higher revenue for the auctioneer. Edelman et al. (2007) speculate that the complexity of VCG and the fact that a VCG potentially leads to lower revenues than the GSP, if advertisers fail to adapt simultaneously to such a change, are reasons for the empirical success of GSP (see also Milgrom, 2004, for a critique of the VCG mechanism in another context). Varian (2007) provides some empirical evidence that the prices observed in Google's ad auctions are in line with the locally envy-free Nash equilibria. Using the same underlying equilibrium model, Varian (2009) finds a value/price relation of about 2–2.3 in ad auctions, while (non-truthful) bids are only about 25 percent to 30 percent larger than prices. Fukuda et al. (2010) compare the GSP auction and the VCG mechanism in a laboratory experiment. They implement ad auctions with 5 advertisers and 5 ad slots. Contrary to the assumptions of the underlying theory discussed above, bidders are not informed on the values of the others. In order to test the robustness of results, two different valuation distributions (a “big keyword” and a “normal keyword” distribution) are implemented that differ in the expected revenues for the advertisers. Click-through-rates of the different ad positions were held constant across conditions. The authors find that revenues in both mechanisms are close to the lower bound of the locally envyfree equilibrium discussed above, but slightly higher in the GSP treatment than in VCG. More equilibrium behavior and higher allocative efficiency is observed in the VCG treatment, and, moreover, both of those measures increase over time, indicating learning and convergence. 4.3. Multi-Unit Demand So far, we have assumed that each bidder only demands a single unit. If bidders demand more than one unit, their strategy space becomes more complex and the theoretical results of revenue equivalence no longer hold. Consider a case with 2 bidders, where A demands only one unit and B demands two. If B wants to buy both units, the price per unit that B must pay must be greater than A's value (as to keep A out of the market). But if B engages in “demand reduction” and only asks for one unit, both units will sell at a price of zero. Demand reduction incentives (or supply-reduction in reverse auctions) exist in all multi-unit auctions. The specific auction format, however, may affect the extent to which demand reduction and collusion amongst bidders is successful (see, e.g., Ausubel and Cramton, 2004, Milgrom, 2004, and the references cited therein). One relatively simple example of multi-unit auctions with multi-unit-demand, among many others, are auctions conducted by governments to sell carbon permits. Many of these auctions are conducted via the Internet, increasing transparency, reducing transaction costs, and allowing a broad spectrum of companies (p. 331) to participate. Yet even in such simple environments, there are many important design choices to be made, including pricing rules, information feedback, timing of markets, product design etc. (see Ockenfels, 2009, and the references therein). In the context of the Regional Greenhouse Gas Initiative (RGGI) in some eastern states of the U.S., Holt et al. (2007) 38 Page 16 of 27 Internet Auctions test the performance of a sealed-bid and a clock auction format in a one-vintage multi-unit auction experiment.38 They find no difference in the allocative efficiency of the auctions. However, in an extension of the original experiment allowing for communication between bidders, Burtraw et al. (2009) observe more bidder collusion and therefore lower seller revenues in a clock auction, in which, unlike in the sealed-bid auction, bidders can update their information about the bidding behavior of others during the course of the auction. Mougeot et al. (2009) show that the existence of speculators in the market is effective in curbing collusion, but in turn may hamper the efficiency of the allocation. Porter et al. (2009) run a two-vintage multi-unit auction experiment in the context of the Virginia NOx auctions. The authors find more efficient allocations and higher seller revenues with a clock auction when permit demand is very elastic. Betz et al. (2011) compare the sequential and simultaneous auctioning of two vintages of multiple units of permits under three different auction designs: a sealed-bid auction and a clock auction with or without demand revelation. Consistent with most previous results, auction formats did not differ significantly with respect to allocative efficiency and seller revenues. An unexpected finding of the experiment is that sequentially auctioning the two vintages yields more efficient allocations, higher auction revenues, and better price signals than auctioning the two vintages simultaneously. While this result stands in contrast with the notion that in general simultaneous allocation procedures allow for higher allocative efficiency than sequential ones, it is in line with a tradeoff between auction complexity and efficiency assumed by many practical auction designers (see, e.g., Milgrom, 2004). Further empirical research is needed to test whether this specific result is robust and generalizes from the lab to real-world auctions.39 Another important complication in multi-unit (and multi-object) auctions arises when there are complementarities between the items auctioned. For example, a buyer might wish to purchase either four or two tires, but neither three nor just one. The uncertainty about the final allocation and potential “exposure” creates strategic incentives for bidders to distort their bids. A possible solution is to elicit bids over all possible packages, thereby allowing bidders to express the complementarities in their bundle values. However, the bidding procedures for such auctions are not only complex, but also pose many strategic difficulties. While many of these issues can be and are addressed, surveying the relevant literature would go beyond the scope of this chapter. Fortunately, there is a large literature dealing with these complexities from a theoretical, empirical, experimental and engineering perspective. See Bichler (2011) de Vries and Vohra (2003) and Cramton et al. (2006) for excellent overviews. Katok and Roth (2004), among others, give an insight on some of the laboratory evidence in the context of Internet auctions. (p. 332) 5. Conclusions This chapter gives a selective overview of theoretical, empirical, and experimental research on Internet auctions. This includes behavioral issues such as overbidding, auction fever, the winner's curse, and sniping. From the seller's perspective we review the effects of setting open or secret reserves, using shill-bidding strategies, and offering buy-it-now options. We also discuss the recent emergence of auction platforms that charge bidding fees and perform unique bid auctions. Finally, we also mention challenges that arise if multiple items are sold in a single auction. One conclusion from this research is that a comprehensive understanding of bidding and outcomes requires a good understanding of both institutional and behavioral complexities, and how they affect each other. This, in turn, requires a broad toolkit, that includes economic theory and field data analyses, as well as behavioral economics, psychology, and operations research. It is the complementing evidence from different perspectives and collected under different degrees of control that allows us to derive robust knowledge on Internet auctions. And it is this knowledge that can make a difference in the real-world. In fact, the emergence of Internet-specific auction formats, like advertisement position auctions, proxy bidding, and penny auctions initiated a whole new research literature that already feeds back into the development and design of these auctions. Research on Internet auctions one of those fields in economics with a direct and mutually profitable link between academic research and industry application. References Page 17 of 27 Internet Auctions Amann, E., Leininger, W., 1995. Expected Revenue of All-Pay and First-Price Auctions with Affiliated Signals. Journal of Economics 61(3), pp. 273–279. Amann, E., Leininger, W., 1996. Asymmetric All-Pay Auctions with Incomplete Information: The Two-Player Case. Games and Economic Behavior 14(1), pp. 1–18. Ambrus, A., Burns, J., 2010. Gradual Bidding in eBay-like Auctions. Working Paper, Harvard University. Anwar, S., McMillan, R., Zheng, M., 2006. Bidding Behavior in Competing Auctions: Evidence from eBay. European Economic Review 50(2), pp. 307–322. Ariely, D., Ockenfels, A., Roth, A., 2005. An Experimental Analysis of Ending Rules in Internet Auctions. The RAND Journal of Economics 36(4), pp. 890–907. Ariely, D., Simonson, I., 2003. Buying, Bidding, Playing, or Competing? Value Assessment and Decision Dynamics in Online Auctions. Journal of Consumer Psychology 13(1–2), pp.113–123. Augenblick, N., 2010. Consumer and Producer Behavior in the Market for Penny Auctions: A Theoretical and Empirical Analysis. Working Paper, Haas School of Business, UC Berkeley. Ausubel, L.M., Cramton, P., 2004. Auctioning Many Divisible Goods. Journal of the European Economic Association 2, pp. 480–493. Bajari, P., Hortaçsu, A., 2003. Winner's Curse, Reserve Prices, and Endogenous Entry: Empirical Insights from eBay Auctions. The RAND Journal of Economics 34(2), pp. 329–355. Bajari, P., Hortaçsu, A., 2004. Economic Insights from Internet Auctions. Journal of Economic Literature 42(2), pp. 457–486. (p. 336) Bapna, R., Goes, P., Gupta, A., 2003. Analysis and Design of Business-to-Consumer Online Auctions. Management Science 49(1), pp. 85–101. Barbaro, S., Bracht, B., 2006. Shilling, Squeezing, Sniping: Explaining latebidding in online second-price auctions. Working Paper, University of Mainz. Baye, M.R., Morgan, J., 2004. Price Dispersion in the Lab and on the Internet: Theory and Evidence. The RAND Journal of Economics 35(3), pp. 449–466. Bazerman, M. H., Samuelson, W. F., 1983. I Won the Auction but Don’t Want the Prize. Journal of Conflict Resolution 27(4), pp. 618–634. Betz, R., Greiner, B., Schweitzer, S., Seifert, S., 2011. Auction Format and Auction Sequence in Multi-item Multi-unit Auctions—An experimental comparison. Working Paper, University of New South Wales. Bichler, M., 2011. Combinatorial Auctions: Complexity and Algorithms. In: Cochran, J.J. (ed.), Wiley Encyclopedia of Operations Research and Management Science. John Wiley & Sons Ltd, Hoboken, N.J. Bolton, G.E., Ockenfels, A., 2000. ERC: A Theory of Equity, Reciprocity, and Competition. American Economic Review 90(1), pp. 166–193. Bolton, G. E., Ockenfels, A., 2010. Does Laboratory Trading Mirror Behavior in Real World Markets? Fair Bargaining and Competitive Bidding on eBay. Working Paper, University of Cologne. Borle, S., Boatwright, P., Kadane, J. B., 2006. The Timing of Bid Placement and Extent of Multiple Bidding: An Empirical Investigation Using eBay Online Auctions. Statistical Science 21(2), pp. 194–205. Brynjolfsson, E., M. D. Smith. 2000. Frictionless commerce? A comparison of Internet and conventional retailers. Management Science 46(4), pp. 563–585. Budish, E. B., Takeyama, L. N., 2001. Buy Prices in Online Auctions: Irrationality on the Internet? Economic Letters, 72(3), pp. 325–333. Page 18 of 27 Internet Auctions Burtraw, D., Goeree, J., Holt, C., Myers, E., Palmer, K., Shobe, W., 2009. Collusion in Auctions for Emission Permits. Journal of Policy Analysis and Management 28(4), pp. 672–691. Casari, M., Ham, J. C., Kagel, J. H., 2007. Selection Bias, Demographic Effects and Ability Effects in Common Value Auction Experiments. American Economic Review 97(4), pp.1278–1304. Cassady, R., 1967. Auctions and Auctioneering. University of California Press, Berkeley. Chakraborty, I., Kosmopoulou, G., 2004. Auctions with shill bidding. Economic Theory 24(2), pp. 271–287. Charness, G., Levin, D., 2009. The Origin of the Winner's Curse: A Laboratory Study. American Economic Journal: Microeconomics 1(1), pp. 207–236. Cooper, D. J., Fang, H., 2008. Understanding Overbidding In Second Price Auctions: An Experimental Study. Economic Journal 118, pp. 1572–1595. Cox, J. C., Roberson, B., Smith, V. L., 1982. Theory and Behavior of Single Object Auctions. In: V. L. Smith (ed.), Research in Experimental Economics, Vol. 2. JAI Press, Greenwich, pp. 1–43. Cox, J. C., Smith, V. L., Walker, J. M., 1985. Experimental Development of Sealed-Bid Auction Theory; Calibrating Controls for Risk Aversion. American Economic Review 75(2), pp. 160–165. Cramton, P., Ausubel, L. M., Filiz-Ozbay, E., Higgins, N., Ozbay, E., Stocking, A., 2009. Common-Value Auctions with Liquidity Needs: An Experimental Test of a Troubled Assets Reverse Auction. Working Paper, University of Maryland. (p. 337) Cramton, P., Shoham, Y., Steinberg, R. (eds.), 2006. Combinatorial Auctions. Cambridge and London: MIT Press. de Vries, S., Vohra, R., 2003. Combinatorial Auctions: A Survey. INFORMS Journal on Computing 5(3), pp. 284–309. Dennehy, M., 2000. eBay adds “buy-it-now” feature. AuctionWatch.com. Dewally, M., Ederington, L. H., 2004. What Attracts Bidders to Online Auctions and What is Their Incremental Price Impact? Working Paper. Duffy, J., Ünver, M. U., 2008. Internet Auctions with Artificial Adaptive Agents: A Study on Market Design. Journal of Economic Behavior & Organization 67(2), pp. 394–417. Durham, Y., Roelofs, M. R., Standifird, S. S., 2004. eBay's Buy-It-Now Function: Who, When, and How. Topics in Economic Analysis and Policy 4(1), Article 28. Edelman, B., Ostrovsky, M., Schwarz, M., 2007. Internet Advertising and the Generalized Second-Price Auction: Selling Billions of Dollars Worth of Keywords. American Economic Review 97(1), pp. 242–259. Eichberger, J., Vinogradov, D., 2008. Least Unmatched Price Auctions: A First Approach. Working Paper 0471, University of Heidelberg. Ely, J. C., Hossain, T., 2009. Sniping and Squatting in Auction Markets. American Economic Journal: Microeconomics 1(2), pp. 68–94. Elyakime, B., Laffont, J. J., Loisel, P., Vuong, Q., 1994. First-Price Sealed-Bid Auctions with Secret Reservation Prices. Annales d’Économie et de Statistique 34, pp. 115–141. Engelberg, J., Williams, J. , 2009. eBay's Proxy System: A License to Shill. Journal of Economic Behavior and Organization 72(1), pp. 509–526. Engelbrecht-Wiggans, R., Katok, E., 2007. Regret in Auctions: Theory and Evidence. Economic Theory 33(1), pp. 81–101. Engelbrecht-Wiggans, R., Katok, E., 2009. A Direct Test of Risk Aversion and Regret in First Price Sealed-Bid Auctions. Decision Analysis 6(2), pp. 75–86. Page 19 of 27 Internet Auctions Fehr, E., Schmidt, K. M., 1999. A Theory Of Fairness, Competition, and Cooperation. The Quarterly Journal of Economics 114(3), pp. 817–868. Filiz-Ozbay, E., Ozbay, E. Y., 2007. Auctions with Anticipated Regret: Theory and Experiment. American Economic Review 97(4), pp. 1407–1418. Fudenberg, D., Tirole, J., 1986. A Theory of Exit in Duopoly. Econometrica 54(4), pp. 943–960. Fukuda, E., Kamijo, Y., Takeuchi, A., Masui, M., Funaki, Y., 2010. Theoretical and experimental investigation of performance of keyword auction mechanisms. GLOPE II Working Paper No. 33. Füllbrunn, S., Sadrieh, A., 2012. Sudden Termination Auctions—An Experimental Study. Journal of Economics and Management Strategy 21(2), pp. 519–540. Gallice, A., 2009. Lowest Unique Bid Auctions with Signals, Working Paper, Carlo Alberto Notebooks 112. Garratt, R., Walker, M., Wooders, J., 2008. Behavior in Second-Price Auctions by Highly Experienced eBay Buyers and Sellers. Working Paper, University of Arizona. Goeree, J., Offerman, T., 2003. Competitive Bidding in Auctions with Private and Common Values. Economic Journal 113, pp. 598–613. Gray, S., Reiley, D., 2007. Measuring the Benefits to Sniping on eBay: Evidence from a Field Experiment. Working Paper, University of Arizona. Grebe, T., Ivanova-Stenzel, R., Kröger, S., 2010. Buy-It-Now Prices in eBay Auctions—The Field in the Lab. Working Paper 294, SFB/TR 15 Governance and the Efficiency of Economic Systems. (p. 338) Greiner, B., Ockenfels, A., 2012. Bidding in Multi-Unit eBay Auctions: A Controlled Field Experiment. Mimeo. Grosskopf, B., Bereby-Meyer, Y., Bazerman, M., 2007. On the Robustness of the Winner's Curse Phenomenon. Theory and Decision 63(4), pp. 389–418. Haile, P. A., 2000. Partial Pooling at the Reserve Price in Auctions with Resale Opportunities. Games and Economic Behavior 33(2), pp. 231–248. Harrison, G. W., 1990. Risk Attitudes in First-Price Auction Experiments: A Bayesian Analysis. Review of Economics and Statistics 72(3), pp. 541–546. Hendricks, K., Onur, I., Wiseman, T., 2008. Last-Minute Bidding in Sequential Auctions with Unobserved, Stochastic Entry. Working Paper, University of Texas at Austin. Heyman, J. E., Orhun, Y., Ariely, D., 2004. Auction Fever: The Effects of Opponents and Quasi-Endowment on Product Valuations. Journal of Interactive Marketing 18(4), pp. 7–21. Hidvégi, Z., Wang, W., Whinston, A. B., 2006. Buy-Price English Auction. Journal of Economic Theory 129(1), pp. 31–56. Holt, C., Shobe, W., Burtraw, D., Palmer, K., Goeree, J., 2007. Auction Design for Selling CO2 Emission Allowances Under the Regional Greenhouse Gas Initiative. Final Report. Ressources for the Future (RFF). Hoppe, T., 2008. An Experimental Analysis of Parallel Multiple Auctions. FEMM Working Paper 08031, University of Magdeburg. Hoppe, T., Sadrieh, A., 2007. An Experimental Assessment of Confederate Reserve Price Bids in Online Auction. FEMM Working Paper 07011, University of Magdeburg. Hossain, T. 2008. Learning by bidding. RAND Journal of Economics 39(2), pp. 509–529. Hossain, T., Morgan, J., 2006. Plus Shipping and Handling: Revenue (Non)Equivalence in Field Experiments on eBay. Advances in Economic Analysis & Policy 6(2), Article 3. Page 20 of 27 Internet Auctions Houser, D., Wooders, J., 2005. Hard and Soft Closes: A Field Experiment on Auction Closing Rules In: R. Zwick and A. Rapoport (eds.), Experimental Business Research. Kluwer Academic Publishers, pp. 123–131. Isaac, R.M., Walker, J.M., 1985. Information and conspiracy in sealed bid auctions. Journal of Economic Behavior & Organization 6, pp. 139–159. Izmalkov, S., 2004. Shill bidding and optimal auctions. Working Paper, Massachusetts Institute of Technology. Jin, G. Z., Kato, A., 2006. Price, Quality, and Reputation: Evidence from an Online Field Experiment. The RAND Journal of Economics 378(4), pp. 983–1004. Kagel, J., 1995. Auctions: A Survey of Experimental Research. In: J. Kagel and A. Roth (eds.), The Handbook of Experimental Economics. Princeton University Press, Princeton, pp.501–585. Kagel, J., Harstad, R., Levin, D., 1987. Information Impact and Allocation Rules in Auctions with Affiliated Private Values: A Laboratory Study. Econometrica 55, pp. 1275–1304. Kagel, J., Levin, D., 1986. The Winner's Curse and Public Information in Common Value Auctions. American Economic Review 76(5), pp. 894–920. Kagel, J., Levin, D., 1993. Independent Private Value Auctions: Bidder Behavior in First-, Second- and Third Price Auctions with Varying Numbers of Bidders. Economic Journal 103(419), pp. 868–879. (p. 339) Kagel, J., Levin, D., 2002. Common Value Auctions and the Winner's Curse. Princeton University Press, Princeton. Katkar, R., Reiley, D. H., 2006. Public Versus Secret Reserve Prices in eBay Auctions: Results from a Pokémon Field Experiment. Advances in Economic Analysis and Policy, Volume 6(2), Article 7. Katok, E., Kwasnica, A. M., 2008. Time is Money: The Effect of Clock Speed on Seller's Revenue in Dutch Auctions. Experimental Economics 11(4), pp. 344–357. Katok, E., Roth, A.E., 2004. Auctions of Homogeneous Goods with Increasing Returns: Experimental Comparison of Alternative “Dutch” Auctions. Management Science 50(8), pp.1044–1063. Kauffman, R. J., Wood, C. A., 2005. The Effects of Shilling on Financial Bid Prices in Online Auctions. Electronic Commerce Research and Applications 4(1), pp. 21–34. Kirchkamp, O., Reiß, J. P., Sadrieh, A., 2006. A pure variation of risk in first-price auctions. FEMM Working Paper 06026, University of Magdeburg Kittsteiner, T., Ockenfels, A., 2008. On the Design of Simple Multi-Unit Online Auctions. In: H. Gimpel, N. R. Jennings, G. E. Kersten, A. Ockenfels, C. Weinhardt (eds.), Lecture Notes in Business Information Processing. Springer, Berlin, pp. 68–71. Klemperer, P., 1999. Auction Theory: A Guide to the Literature. Journal of Economic Surveys 13, pp. 227–286. Kosmopoulou, G., de Silva, D. G., 2007. The effect of shill bidding upon prices: Experimental evidence. International Journal of Industrial Organization 25(2), pp. 291–313. Krishna, V., 2002. Auction Theory. Academic Press, New York. Ku, G., Malhotra, D., Murnighan, J. K., 2005. Towards A Competitive Arousal Model of Decison-Making: A Study of Auction Fever in Live and Internet Auctions. Organizational Behavior and Human Decision Processes 96(2), pp. 89– 103. Lee, Y. H., Malmendier, U., 2011. The Bidder's Curse. American Economic Review 101(2), pp. 749–787. Levin, D., Smith, J. L. , 1996. Optimal Reservation Prices in Auctions. Economic Journal 106(438), pp. 1271–1282. Li, H., Tan, G., 2000. Hidden Reserve Prices with Risk-Averse Bidders. Working Paper, University of British Columbia. Page 21 of 27 Internet Auctions Lind, B., Plott, C. R., 1991. The Winner's Curse: Experiments with Buyers and with Sellers. American Economic Review 81(1), pp. 335–346. Lucking-Reiley, D., 1999. Using Field Experiments to Test Equivalence Between Auction Formats: Magic on the Internet. American Economic Review 89(5), pp. 1063–1080. Lucking-Reiley, D., 2000a. Auctions on the Internet: What's Being Auctioned, and How? Journal of Industrial Economics 48(3), pp. 227–252. Lucking-Reiley, D., 2000b. Vickrey Auctions in Practice: From Ninetheenth Century Philately to Twenty-FirstCentrury E-Commerce. Journal of Economic Perspectives 14(3), pp. 183–193. Lucking-Reiley, D., Bryan, D., Prasd, N., Reeves, D., 2007. Pennies from Ebay: The Determinants of Price in Online Auctions. Journal of Industrial Economics 55(2), pp. 223–233. Maskin, E., Riley, J. G., 1984. Optimal Auctions with Risk Averse Buyers. Econometrica 52(6), pp. 1473–1518. (p. 340) Mathews, T., Katzman, B., 2006. The Role of Varying Risk Attitudes in an Auction with a Buyout Option. Economic Theory 27(3), pp. 597–613. Matthews, S., 1987. Comparing Auctions for Risk Averse Buyers: A Buyer's Point of View. Econometrica 55(3), pp. 633–646. McAfee, R. P., 1993. Mechanism Design by Competing Sellers. Econometrica 61(6), pp.1281–1312. McAfee, R. P., McMillan, J., 1987. Auctions with Entry. Economic Letters 23(4), pp. 343–347. McCart, J. A., Kayhan, V. O., Bhattacherjee, A., 2009. Cross-Bidding in Simultaneous Online Auctions. Communications of the ACM 52(5), pp. 131–134. Milgrom, P., 2004. Putting Auction Theory to Work. Cambridge University Press, Cambridge. Milgrom, P., Weber, R. J., 1982. A Theory of Auctions and Competitive Bidding. Econometrica 50(5), pp. 1089–1122. Morgan, J., Steiglitz, K., Reis, G., 2003. The Spite Motive and Equilibrium Behavior in Auctions. Contributions to Economic Analysis & Policy 2(1), Article 5. Mougeot, M., Naegelen, F., Pelloux, B., Rullière, J.-L., 2009. Breaking Collusion in Auctions Through Speculation: An Experiment on CO2 Emission Permit Market. Working paper, GATE CNRS. Myerson, R. B., 1981. Optimal Auction Design. Mathematics of Operations Research 6(1), pp. 58–73. Myerson, R. B. , 1998. Population Uncertainty and Poisson Games. International Journal of Game Theory 27, pp. 375–392. Nagareda, T., 2003. Announced Reserve Prices, Secret Reserve Prices, and Winner's Curse. Mimeo. Nekipelov, D. 2007. Entry deterrence and learning prevention on eBay. Working paper, Duke University, Durham, N.C. Ockenfels, A., 2009. Empfehlungen fuer das Auktionsdesign fuer Emissionsberechtigungen. [Recommendations for the design of emission allowances auctions.] Zeitschrift fuer Energiewirtschaft (2), pp. 105–114. Ockenfels, A., Reiley, D., Sadrieh, A., 2006. Online Auctions. In: Hendershott, T. J. (ed.), Handbooks in Information Systems I, Handbook on Economics and Information Systems, pp.571–628. Ockenfels, A., Roth, A., 2006. Late and Multiple Bidding in Second Price Internet Auctions: Theory and Evidence Concerning Different Rules for Ending an Auction. Games and Economic Behavior 55(2), pp. 297–320. Ockenfels, A., Roth, A., 2010. Ending Rules in Internet Auctions: Design and Behavior. Working paper. Page 22 of 27 Internet Auctions Ockenfels, A., Selten, R., 2005. Impulse Balance Equilbrium and Feedback in First Price Auctions. Games and Economic Behavior 51(1), pp. 155–179. Ockenfels, A., Tillmann, P., Wambach, A., 2010. English All-Pay Auctions—An Empirical Investigation. Mimeo, University of Cologne. Östling, R., Wang, J. T., Chou, E., Camerer, C. F., 2011. Testing Game Theory in the Field: Swedish LUPI Lottery Games. American Economic Journal: Microeconomics 3(3), pp. 1–33. Peeters, R., Strobel, M., Vermeulen, D., Walzl, M., 2007. The impact of the irrelevant—Temporary buy-options and bidding behavior in online auctions. Working Paper RM/07/027, Maastricht University. (p. 341) Peters, M., Severinov, S., 2006. Internet Auctions with Many Traders. Journal of Economic Theory 130(1), pp. 220– 245. Popkowski Leszczyc, P. T. L. , Qiu, C., He, Y., 2009. Empirical Testing of the Reference-Price Effect of Buy-Now Prices in Internet Auctions. Journal of Retailing 85(2), pp. 211–221. Porter, D., Rassenti, S., Shobe, W., Smith, V., Winn, A., 2009. The design, testing and implementation of Virginia's NOx allowance auction. Journal of Economic Behavior & Organization 69(2), pp.190–200. Rapoport, A., Otsubo, H., Kim, B., Stein, W.E., 2009. Unique Bid Auction Games. Jena Economic Research Papers 2009–005. Rasmusen, E., 2007. Getting Carried Away in Auctions as Imperfect Value Discovery. Mimeo. Kelley School of Business, Indiana University. Raviv, Y., Virag, G., 2009. Gambling by auctions. International Journal of Industrial Organization 27, pp. 369–378. Reiley, D. H., 2006. Field Experiments on the Effects of Reserve Prices in Auctions: More Magic on the Internet. RAND Journal of Economics 37(1), pp. 195–211. Reynolds, S. S. and Wooders, J., 2009. Auctions With a Buy Price. Economic Theory 38(1), pp. 9–39. Riley, J. G., 1980. Strong Evolutionary Equilibrium and the War of Attrition. Journal of Theoretical Biology 82(3), pp. 383–400. Riley, J. G., Samuelson, W. F., 1981. Optimal Auctions. American Economic Review 71(3), pp. 381–392. Rose, S. L., Kagel, J., 2009. Almost Common Value Auctions: An Experiment. Journal of Economics & Management Strategy 17(4), pp. 1041–1058. Rosenkranz, S., Schmitz, P. W., 2007. Reserve Prices in Auctions as Reference Points. Economic Journal 117(520), pp. 637–653. Roth, A., Ockenfels, A., 2002. Last-Minute Bidding and the Rules for Ending Second-Price Auctions: Evidence from eBay and Amazon Auctions on the Internet. American Economic Review 92(4), pp. 1093–1103. Samuelson, W. F., 1985. Competitive Bidding with Entry Costs. Economic Letters 17(1–2), pp. 53–57. Seifert, S., 2006. Posted Price Offers in Internet Auction Markets. Springer, Berlin, Heidelberg, New York. Shahriar, Q., Wooders, J., 2011. An Experimental Study of Auctions with a Buy Price Under Private and Common Values. Games and Economic Behavior 72(2), pp. 558–573. Simonsohn, U., 2010. eBay's Crowded Evenings: Competition Neglect in Market Entry Decisions. Management Science 56(7), pp. 1060–1073. Simonsohn, U., Ariely, D., 2008. When rational sellers face nonrational consumers: Evidence from herding on eBay. Management Science 54(9), pp. 1624–1637. Page 23 of 27 Internet Auctions Song, J., Baker, J., 2007. An Integrated Model Exploring Sellers’ Strategies in eBay Auctions. Electronic Commerce Research 7(2), pp. 165–187. Stern, B. B., Stafford, M. R., 2006. Individual and Social Determinants of Winning Bids in Online Auctions. Journal of Consumer Behaviour 5(1), pp. 43–55. Thaler, R. H., 1980. Toward a Positive Theory of Consumer Choice. Journal of Economic Behavior & Organization 1(1), pp. 39–60. Thaler, R. H., 1988. Anomalies: The Winner's Curse. Journal of Economic Perspectives 2(1), pp. 191–202. (p. 342) Varian, H., 2007. Position Auctions. International Journal of Industrial Organization 25(6), pp. 1163–1178. Varian, H., 2008. Quality scores and ad auctions. Available at: http://googleblog.blogspot.com/2008/10/quality-scores-and-ad-auctions.html Varian, H., 2009. Online Ad Auctions. American Economic Review 99(2), pp. 430–434. Vickrey, W., 1961. Counterspeculation Auction and Competitive Sealed Tenders. Journal of Finance 16(1), pp. 8– 37. Vincent, D. R., 1995. Bidding Off the Wall: Why Reserve Prices May Be Kept Secret. Economic Theory 65(2), pp. 575–584. Walker, J. M., Smith, V. L., Cox, J. C., 1990. Inducing Risk Neutral Preferences: An Examination in a Controlled Market Environment. Journal of Risk and Uncertainty 3, pp. 5–24. Wilcox, R. T., 2000. Experts and Amateurs: The Role of Experience in Internet Auctions. Marketing Letters 11(4), pp. 363–374. Wintr, L., 2008. Some Evidence on Late Bidding in eBay Auctions. Economic Inquiry 46(3), pp. 369–379. Yin, P.-L., 2009. Information Dispersion and Auction Prices. Mimeo. Sloan School of Management, Massachusetts Institute of Technology. Zeithammer, R., Adams, C., 2010. The Sealed-Bid Abstraction in Online Auctions. Marketing Science 29(6), pp. 964– 987. Notes: (1.) Cassady (1967) surveys the history of auctions. (2.) Auction fees are also much lower on Internet auction platforms than in traditional offline-auctions (5–10 percent as compared to about 20 percent, respectively; Lucking-Reiley, 2000a). (3.) Lucking-Reiley (1999) explores further advantages of Internet auctions. (4.) Bajari and Hortasçu (2004), Lucking-Reiley (2000a), and Ockenfels et al. (2006) also provide surveys of this research. (5.) The research on reputation systems is closely related to the research on Internet auctions because of the important role they play in enabling trade on these platforms. Such systems have natural ties with the design of auctions themselves, as they influence the seller-buyer interaction. We do not discuss reputation systems in this chapter, as they are dealt with in detail in chapter 13 of this book. (6.) See Klemperer (1999), Krishna (2002), and Milgrom (2004), among others, for more comprehensive and formal treatments of auction theory. (7.) Outside economics, the term “Dutch auction” is often used differently, e.g., for multi-unit auctions. Page 24 of 27 Internet Auctions (8.) The affiliated values model was introduced by Milgrom and Weber (1982). A combined value model can be found, e.g., in Goeree and Offerman (2003). (9.) In the literature, this case is often also referred to as the “i.i.d.” model, because values are “independently and identically distributed.” (10.) The term “standard auction” refers to single-item auctions with symmetric private values, in which the highest bid wins the item. (11.) As a result, if bidders are not risk-neutral, some auction formats are more attractive than others, where the ranking is different for sellers and buyers (Maskin and Riley, 1984; Matthews, 1987). (12.) See Kagel (1995) for an extensive survey on auction experiments. (13.) For example, Kagel and Levin (1993) observe overbidding in sealed-bid auctions, and Katok and Kwasnica (2008) in open decreasing-price auctions. (14.) Overbidding is sometimes also observed in second-price auctions, where risk aversion should not play a role, and third-price auctions, where risk aversion even predicts underbidding (Kagel and Levin, 1993). Harrison (1990), Cox et al. (1982), and Walker et al. (1990) find substantial overbidding, although they control for risk aversion through either directly measuring it or by using experimental techniques to induce risk-neutral preferences. Kagel (1995) provides a detailed discussion of the literature on risk aversion in auctions. (15.) Bolton and Ockenfels (2010) conducted a controlled field experiment on eBay (inducing private values, etc.), and examined to what extent both social and competitive laboratory behavior are robust to institutionally complex Internet markets with experienced traders. Consistent with behavioral economics models of social comparison (e.g., Fehr and Schmidt, 1999; Bolton and Ockenfels, 2000), they identify an important role of fairness in one-toone eBay interactions, but not in competitive one-to-many eBay auctions. This suggests that social concerns are at work in Internet trading; yet the study cannot (and was not designed to) reveal a role of social utility for overbidding. (16.) The first database consists of 167 auctions of a particular board game. The second database consists of 1,926 auctions of very heterogeneous items. (17.) A distinct but related phenomenon is herding on Internet auctions; see Simonsohn and Ariely (2008) and Simonsohn (2010). (18.) A full picture must also take into account the literature on price dispersion on the Internet (see, e.g., Baye and Morgan, 2004, and Brynjolfsson and Smith, 2000, and the references therein). (19.) In contrast, laboratory experiments by Kagel and Levin (1986) suggest that common value auctions with larger numbers of bidders produce even more aggressive bidding. (20.) As a complementary result, Jin and Kato (2006) find that sellers with better reputation profile are able to obtain higher prices. These sellers are also less likely to default, but conditional on delivery, these sellers are not less likely to make false claims or deliver low quality. (21.) Similarly, Bajari and Hortaçsu (2003) observe that half of the winning bids in their sample of eBay auctions arrive in the last 1.7 percent (~73 minutes) of the auction time, and 25 percent of the winning bids were submitted in the last 0.2 percent of the time period. (22.) Some authors also point out that late bidding maybe due to bidders’ learning processes. Duffy and Ünver (2008), for example, show that finite automata playing the auctions repeatedly and updating their strategies via a genetic algorithm also exhibit bidding behavior that is akin to the observed behavior in the field. (23.) This result is supported by the experimental study of Füllbrunn and Sadrieh (2006), who observe substantial late bidding in hard-close and candle auctions with a 100 percent probability of bid arrival. Interestingly, they show that in candle auctions, i.e. auctions that have a stochastic termination time, late bidding always sets in the first period with a positive termination probability. This indicates that bidders are fully aware of the strategic effect of the Page 25 of 27 Internet Auctions stochastic termination rule, but choose to delay serious bidding as long as it is possible without the threat of a sudden “surprise” ending. (24.) Strict incremental bidding equilibria, however, are neither observed in the field (McCart et al. 2009) nor in laboratory experiments with competing auctions (Hoppe, 2008), even though both incremental and late bidding are present. (25.) E.g., Cox et al. (1982) find higher revenues in the first-price sealed-bid auction compared to a Dutch auction. Kagel et al. (1987) observe higher revenues for the second-price auction compared to the English auction, though the experimental results for the English auction are in accordance with the theoretical predictions. (26.) The finding that reserve prices increase final auction prices conditional on sale is replicated in field data analyses of Lucking-Reiley et al. (2007) and Stern and Stafford (2006). (27.) Sometimes, however, the auction rules may actually help the prohibited shill bidding behavior. Engelberg and Williams (2009) report a particular feature of the incrementrule on eBay that makes dynamic shill bidding almost fool-proof. When a new bid arrives, eBay generally increases the current price by at least a full increment. If, however, the proxy bid of the highest bidder is less than one increment above the new bid, the increment is equal to the actual distance between the two highest bids. This feature allows sellers to submit shill bids incrementally until the price increases by less than an increment, i.e. the highest proxy bid is reached. This “discover and stop” strategy obviously enables sellers to extract almost the entire buyer's rent. (28.) See, however, Bolton and Ockenfels (2010), who observed shill bidding of their subjects in a controlled field experiment. (29.) Bidding rights can be prepurchased: swoopo.com uses its own auction mechanism to sell bid packages. Swoopo is an international platform, presenting the same auctions simultaneously in different countries, thereby attracting a large number and constant stream of bidders. Bid increments in other countries are €0.01 or £0.01, for example, and bidding fees are €0.50 or £0.50, respectively. (30.) Swoopo is the most successful platform of penny auctions, with about 200 auctions per day (Ockenfels et al., 2010; Augenblick, 2010) and 80,000 conducted auctions in 2008 (self-reported). Currently, the competitors of swoopo.com, like bidstick, rockybid, gobid, bidray and zoozle, only make 7 percent of the industry profits, with the remaining 93 percent captured by swoopo (Augenblick, 2010). (31.) Unique bid auctions seem to be particularly popular for standard consumer goods. There also exists a variant in which the highest unique bid wins. (32.) Note that not bidding also has a positive probability in the mixed equilibrium strategy. (33.) Östling et al. (2011) also find declining choice probabilities and bidder heterogeneity in a very closely related game. They study a LUPI (lowest unique positive integer) lottery, in which participants pay for each submitted bid (i.e., all-pay setting), but the winner receives a money prize without having to pay his bid (i.e., bids are lottery tickets and there is no private value complication). Using the concept of Poisson-Nash equilibrium (Myerson, 1998), Östling et al. (2011) demonstrate that their field and lab data are well in line with the fully mixed symmetric equilibrium with declining choice probabilities, especially if they extend their theoretical analysis to encompass behavioral aspects such as learning, depth of reasoning, and choice errors. (34.) The same generalization holds true in the case of multiple single-unit auctions conducted simultaneously or sequentially: Any auction winner who buys at a price higher than the lowest price among all n auctions would have better chosen a different bidding strategy. Thus, in equilibrium, all n auctions should yield the same price, the n bidders with the highest values should win the n units, and all auction winners should pay a price in the size of the n+1th-highest value among the bidders. (35.) Bapna et al. (2003) take an empirical optimization approach to analyze multi-unit auctions and find the bid increment to have a pivotal role for the seller's revenue. (36.) Since the search engine only earns income, when there are clicks on the search term, high per-click prices with very little traffic can easily be dominated by low per-click prices with much more traffic. Page 26 of 27 Internet Auctions (37.) As long as the click-through-rates (CTR) of all ads are almost equal, the GSP will do a relatively good job in allocating the advertising space efficiently. If, however, an ad with a high bid attracts a very little traffic (i.e., has a small CTR), then it generate less revenue than a high-CTR ad with a lower bid. (38.) In carbon permit auctions, a large number of perfectly substitutable identical items (permits) are sold, often in different “vintages” (batches of permits valid over different predefined time horizons). (39.) An underexplored topic in permit auction research is the effect of the existence of secondary markets, i.e. the possibility of resale (Haile, 2000), which in theory turns a private-value auction into one with common (resale) value. Ben Greiner Ben Greiner is Lecturer at the School of Economics at the University of New South Wales. Axel Ockenfels Axel Ockenfels is Professor and Chairperson of the Faculty of Management, Economics, and Social Sciences at the University of Cologne. He models bounded rationality and social preferences and does applied work behavioral and market design economics. Abdolkarim Sadrieh Abdolkarim Sadrieh is Professor of Economics and Management at the University of Magdeburg. Page 27 of 27 Reputation on the Internet Oxford Handbooks Online Reputation on the Internet Luis Cabral The Oxford Handbook of the Digital Economy Edited by Martin Peitz and Joel Waldfogel Print Publication Date: Aug 2012 Online Publication Date: Nov 2012 Subject: Economics and Finance, Economic Development DOI: 10.1093/oxfordhb/9780195397840.013.0013 Abstract and Keywords This article reports the recent, mostly empirical work on reputation on the Internet, with a particular focus on eBay's reputation system. It specifically outlines some of the main economics issues regarding online reputation. Most of the economics literature has focused on eBay despite the great variety of online reputation mechanisms. Buyers are willing to pay more for items sold by sellers with a good reputation, and this is reflected in actual sale prices and sales rates. It is also noted that reputation matters for buyers and for sellers. Regarding game theory's contribution, it is significant to understand the precise nature of agent reputation in online platforms. Online markets are important, in monetary terms and otherwise; and they are bound to become even more so in the future. Keywords: online reputation, Internet, eBay, economics, buyers, sellers, online markets 1. Introduction: What is Special About Reputation on the Internet? Economists have long studied the phenomenon of reputation, broadly defined as what agents (e.g., buyers) believe or expect from other agents (e.g., sellers). In a recent (partial) survey of the literature (Cabral, 2005), I listed several dozen contributions to this literature. So it is only fair to ask: What is special about reputation on the Internet? Why should a chapter like this be worth reading? There are several reasons that reputation on the Internet is a separate phenomenon in its own right, and one worth studying. First, the growth of the Internet has been accompanied by the growth of formal, systematic review and feedback systems (both of online market agents and of offline market agents who are rated online). By contrast, the traditional research on reputation was typically motivated by real-world problems where only “soft” information was present (for example, the reputation of an incumbent local monopolist such as American Airlines for being “tough” in dealing with entrants into its market). The second reason that reputation on the Internet forms a separate research area is that formal online reputation systems generate a wealth of information hitherto unavailable to the researcher. As a result, the economic analysis of online reputation is primarily empirical, whereas the previous literature on reputation was primarily of a theoretical nature. A third reason that the present endeavor is worthwhile is that online markets are important and increasingly so. Specifically, consider the case of eBay, one of (p. 344) the main online markets and the most studied in the online reputation literature. In 2004, more than $34.1 billion were transacted on eBay by more than one hundred million users. Although much of the research has been focused on eBay, other online markets are also growing very rapidly. China's taobao.com, for example, has more than 200 million registered users as of 2010. Finally, reputation on the Internet is potentially quite important because there is fraud on Internet commerce. Here Page 1 of 9 Reputation on the Internet are three examples, all from 2003: (1) an eBay seller located in Utah sold about 1,000 laptop computers (for about $1,000 each) that were never delivered; (2) a buyer purchased a “new” electrical motor which turn out to be “quite used”; (3) a seller saw her transaction payment reversed upon receiving a “Notice of Transaction Review” from PayPal stating that the funds used to pay for the transaction came “from an account with reports of fraudulent bank account use.”1 While there are various mechanisms to deal with fraud, reputation is one of the best candidates—and arguably one of the more effective ones. In what follows, I summarize some of the main economics issues regarding online reputation. My purpose is not to provide a systematic, comprehensive survey of the literature, which has increased exponentially in recent years. Rather, I point to what I think are the main problems, the main stylized facts, and the main areas for future research. 2. Online Reputation Systems: Ebay and Others As I mentioned earlier, one of the distinguishing features of online reputation is the existence of formal feedback and reputation mechanisms; within these the eBay feedback system is particularly important, both for the dollar value that it represents and for the amount of research it has induced. A brief description of eBay and its feedback system is therefore in order. Since its launch in 1995, eBay has become the dominant online auction site, with millions of items changing hands every day. eBay does not deliver goods: it acts purely as an intermediary through which sellers can post auctions and buyers bid. eBay obtains its revenue from seller fees, based on a complex schedule that includes fees for starting an auction and fees on successfully completed auctions. Most important, to enable reputation to regulate trade, eBay uses an innovative feedback system.2 Originally, after the completion of an auction, eBay allowed both the buyer and the seller to give the other party a grade of +1 (positive), 0 (neutral), or –1 (negative), along with any textual comments. There have been several changes on eBay regarding how these ratings can be given by the users. For example, since 1999 each grade or comment has to be linked to a particular transaction on eBay. Further changes were introduced in 2008, when eBay revoked the ability (p. 345) of the seller to give buyers a negative or neutral grade (sellers could choose to leave positive or no feedback for buyers). However, the core of the feedback system has remained the same.3 Based on the feedback provided by agents (buyers and sellers), eBay displays several aggregates corresponding to a seller's reputation, including (1) the difference between the number of positive and negative feedback ratings; (2) the percentage of positive feedback ratings; (3) the date when the seller registered with eBay; and (4) a summary of the most recent feedback received by the seller.4 Finally, eBay provides a complete record of the comments received by each seller, starting with the most recent ones. All of the information regarding each seller is publicly available; in particular it is available to any potential buyer. While this chapter focuses primarily on literature regarding the eBay platform, eBay is by no means the only online feedback and reputation system. Amazon.com offers a system of customer reviews whereby buyers can rate both the product itself and, if buying from a seller other than Amazon, the seller. Although Amazon's seller review system is quite similar to that of eBay, its product review system is somewhat more complex, as reviewers can rate other reviews. Unlike eBay, Amazon's seller reputation system is one-sided: buyers review sellers but sellers do not review buyers (eBay still offers sellers the option to rate buyers positively). Of course eBay is not the only twosided system. For example, at couchsurfing.net, where travelers find free places to stay while traveling, both hosts and travelers can rate each other. Despite the great variety of online reputation mechanisms, most of the economics literature has focused on eBay. This is partly justified by the economic significance of eBay, the size of which dwarfs almost all other online trade platforms, and by the fact that a considerable amount of data is available for eBay. Accordingly, most of the evidence presented in this chapter corresponds to eBay. 3. Do Buyers Care About Online Reputations? The fact that there exists a feedback reputation system in a given online market does not imply that such system Page 2 of 9 Reputation on the Internet matters, that is, that it has any practical significance. In principle, it is possible that agents (buyers and sellers) ignore the reputation levels generated by the system. If that were the case, there would be little incentive to provide feedback, which in turn would justify the agents’ not caring for the system in the first place. More generally, many if not most games of information transmission (such as feedback and rating systems) admit “babbling equilibria,” that is, equilibria where agents provide feedback and ratings in a random way (or simply don’t provide any feedback), and, consistently, agents ignore the information generated by the system. (p. 346) For this reason, an important preliminary question in dealing with online reputation systems is, Does the system matter at all? In other words, do the reputations derived from formal feedback systems have any bite? One first step in answering this question is to determine whether buyers’ actions (whether to purchase; how much to bid, in case of an auction; and so forth) depend on seller's reputation. At the most basic level, we would expect a better seller reputation to influence the price paid for an otherwise identical object. Many studies attempt to estimate the following equation: sale price as a dependent variable and seller reputation as an independent variable (along with other independent variables). Alternative left-hand side variables include the number of bids (in the case of an auction) or the likelihood the item in question is sold. These studies typically find a weak relation between reputation and price. However, as is frequently the case with cross-section regressions, there are several problems one must take into account. First, unobserved heterogeneity across sellers and sold items may introduce noise in the estimates, possibly reducing the estimates’ statistically significance. Conversely, as first pointed out by Resnick et al. (2003), several unobservable confounding factors may lead to correlations for which there is no underlying causality relation. For example, sellers with better reputation measures may also be much better at providing accurate and clear descriptions of the items they are selling, which in turn attract more bidders; hence their writing ability, not their reputation, may be underlying cause of the higher prices they receive. (In fact, there is some evidence that spellings mistakes in item listings are correlated with lower sale prices; and that “spelling arbitrage” is a profitable activity—that is, buying items with misspelled listings and selling them with correctly spelled listings.) Such caveats notwithstanding, a series of authors have addressed the basic question of the effect of reputation on sales rates and sales price by taking the crosssection regression approach. The list includes Cabral and Hortacsu (2010), Dewan and Hsu (2004), Eaton (2005), Ederington and Dewally (2003), Houser and Wooders (2005), Kalyanam and McIntyre (2001), Livingston (2005), Lucking-Reiley, Bryan, Prasad, and Reeves (2006), McDonald and Slawson (2002), Melnik and Alm (2002), and Resnick and Zeckhauser (2002).5 For example, Cabral and Hortacsu (2010) find that a 1 percent level increase in the fraction of negative feedback is correlated with a 7.5 percent decrease in price, though the level of statistical significance is relatively low. These results are comparable to other studies, both in terms of coefficient size and in terms of statistical significance. One way to control for seller heterogeneity is to go beyond cross-section regression and estimate the effects of reputation based on panel data. From a practical point of view, creating a panel data of online sellers is much more difficult than creating a cross-section. For some items on eBay, it suffices to collect data for a few days in order obtain hundreds if not thousands of observations from different sellers. By contrast, creating a panel of histories of a given set of sellers takes time (or money, if one is to purchase an existing data set).6 Cabral and Hortacsu (2010) propose a strategy for studying eBay seller reputation. At any moment in time, eBay posts data on a seller's complete feedback history. (p. 347) Although there is no information regarding past transactions’ prices, the available data allows for the estimation of some seller reputation effects. Specifically, Cabral and Hortacsu (2010) propose the following working assumptions (and provide statistical evidence that they are reasonable working assumptions): (1) the frequency of buyer feedback is a good proxy for the frequency of actual transactions; (2) the nature of the feedback is a good proxy for the degree of buyer satisfaction. Based on these assumptions, they collect a series of seller feedback histories and estimate the effect of reputation on sales rate. They find that when a seller first receives negative feedback, his weekly sales growth rate drops from a positive 5 percent to a negative 8 percent. (A disadvantage of using seller feedback histories is that one does not obtain price effects, only quantity effects.) As an alternative to panel data, one strategy for controlling for omitted-variable biases is to perform a controlled field experiment. Resnick, Zeckhauser, Swanson, and Lockwood (2006) do precisely that: they perform a series of sales of identical items (collector's postcards) alternatively using a seasoned seller's name and an assumed name Page 3 of 9 Reputation on the Internet with little reputation history. They estimate an 8 percent premium to having 2,000 positive feedbacks and 1 negative over a feedback profile with 10 positive comments and no negatives. In a related research effort, Jin and Kato (2005) assess whether the reputation mechanism is able to combat fraud by purchasing ungraded baseball cards with seller-reported grades, and having them evaluated by the official grading agency. They report that while having a better seller reputation is a positive indicator of honesty, reputation premia or discounts in the market do not fully compensate for expected losses due to seller dishonesty. A related, alternative strategy consists of laboratory experiments. Ba and Pavlou (2002) conduct a laboratory experiment in which subjects are asked to declare their valuations for experimenter generated profiles, and find a positive response to better profiles. As every other research method, laboratory experiments have advantages and disadvantages. On the plus side, they allow the researcher to create a very tightly controlled experiment, changing exactly one parameter at a time; they are the closest method to that of the physical sciences. On the minus side, a major drawback is that the economics laboratory does not necessarily reflect the features of a realworld market. In summary, while different studies and different research methodologies arrive at different numbers, a common feature is that seller reputation does have an effect on buyer behavior: buyers are willing to pay more for items sold by sellers with a good reputation, and this is reflected on actual sale prices and sales rates. As I mentioned already, from a theoretical point of view there could exist an equilibrium where buyers ignore seller reputation (consistently with the belief that feedback is given in a random manner); and seller feedback is given in a random manner (consistently with the fact that seller reputation is ignored by buyers). The empirical evidence seems to reject this theoretical possibility. Finally, I should mention that buyer feedback and seller reputation are not the only instrument to prevent opportunistic behavior on the seller's part. In traditional markets, warranties play an important role in protecting buyers. What (p. 348) role can warranties play in online markets? Roberts (2010) addresses this issue by examining the impact of the introduction of eBay's buyer protection program, a warranty system. He shows that, under the new system, the relation between seller reputation and price becomes “flatter.” This suggests that warranties and seller reputation are (partial) substitute instruments to protect consumers in a situation of asymmetric information in the provision of quality products. 4. Do Sellers Care About Online Reputations? Given that reputations do have a bite, in the sense that buyers care about it, a natural next question is, What do sellers do about it? Specifically, interesting questions include, How do sellers build a reputation? How do sellers react to negative feedback: by increasing or by decreasing effort to provide quality? Do sellers use or squander their reputation by willfully cheating buyers? Cabral and Hortacsu (2010), in the essay introduced earlier, address some of these questions. As previously mentioned, following the first negative feedback received by a seller, the sales growth rate drops from a positive 5 percent to a negative –8 percent. This suggests that buyers care about seller reputation. In addition, Cabral and Hortacsu also show that following the first negative feedback given to the seller, subsequent negative feedback ratings arrive 25 percent more frequently and don’t have nearly as much impact as the first one. This change, they argue, is due to a shift in seller behavior. Intuitively, while a seller's reputation is very high, the incentives to invest on such reputation are also high. By contrast, when a perfect record is stained by a first negative, sellers are less keen on making sure buyer satisfaction is maximized; and as a result negative feedback is received more frequently.7 Cabral and Hortacsu (2010) also find that a typical seller starts his career with a substantially higher fraction of transactions as a buyer relative to later stages of his career as an eBay trader. This suggests that sellers invest in building a reputation as a buyer and then use that reputation as a seller. Moreover, a seller is more likely to exit the lower his reputation is; and just before exiting sellers receive more negative feedback than their lifetime average. Note that the “end of life” evidence is consistent with two different stories. First, it might be that exit is planned by the seller and his reputation is “milked down” during the last transactions. Alternatively, it might be that the seller was hit by some exogenous shock (he was sick for a month and could not process sales during that period), which Page 4 of 9 Reputation on the Internet led to a series of negative feedbacks; and, given the sudden drop in reputation, the seller decides that it is better to exit. Additional evidence is required to choose between these two stories. Anecdotal evidence suggests that the (p. 349) former plays an important role. For example, in their study of sales of baseball cards on eBay Jin and Kato (2005) report that they “encountered two fraudulent sellers who intentionally built up positive ratings, committed a series of defaults, received over 20 complaints, and abandoned the accounts soon afterward” (p. 985). In sum, the evidence suggests that reputation matters not only for buyers but also for sellers. In particular, sellers’ actions too are influenced by reputation considerations. 5. The Feedback Game User feedback is the backbone of online reputations. How and why do feedback mechanisms work? To an economist following the classical homo economicus model, the answer is not obvious. Giving feedback takes time. Moreover, to the extent that feedback can be given by both parties to a transaction, giving feedback may also influence the other party's decision to give feedback (and the nature of such feedback). One must therefore be careful about measuring the costs and benefits of feedback-giving. Formal feedback and review mechanisms induce relatively well-defined extensive form games: each agent must decide, according to some rules, if and when to send messages; and what kind of messages to send. Understanding the equilibrium of these games is an important step towards evaluating the performance of reputation mechanisms, both for online and offline markets. Empirically, one noticeable feature of the eBay feedback mechanism is that there is a very high correlation between the events of a buyer providing feedback to seller and a seller providing feedback to buyer. Jian et al. (2010) argue that eBay sellers use a “reciprocate only” strategy about 20 to 23 percent of the time. Bolton et al. (2009) also provide information that supports the reciprocal nature of feedback giving. For example, they show that the timing of feedback-giving by buyer and seller is highly correlated. The reciprocal nature of feedback also seems consistent with another important stylized fact from the eBay system (and possibly from other two-way feedback systems): the extremely low prevalence of negative feedback. Klein et al. (2006) and Li (2010) present confirmatory evidence. They show that a substantial fraction of the (little) negative feedback that is given takes place very late during the feedback window, presumably in a way that reduces the likelihood of retaliation. Conversely, positive feedback is given very early on, presumably as a way to encourage reciprocal behavior. Given the strategic nature of feedback giving, several proposals have been made to make the process more transparent and closer to truth telling. For example, Bolton et al. (2009) propose that feedback be given simultaneously, thus preventing one party from reacting to the other. They also present evidence from Brazilian (p. 350) MercadoLivre which suggests that such a system induces more “sincere” feedback-giving without diminishing the overall frequency of feedback (one of the concerns with simultaneous feedback). Additional evidence is obtained by considering RentACoder.com, a s